4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
10 #include <asm/switch_to.h>
13 #include "../workqueue_internal.h"
14 #include "../smpboot.h"
16 #define CREATE_TRACE_POINTS
17 #include <trace/events/sched.h>
19 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
21 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
23 * Debugging: various feature bits
25 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
26 * sysctl_sched_features, defined in sched.h, to allow constants propagation
27 * at compile time and compiler optimization based on features default.
29 #define SCHED_FEAT(name, enabled) \
30 (1UL << __SCHED_FEAT_##name) * enabled |
31 const_debug
unsigned int sysctl_sched_features
=
38 * Number of tasks to iterate in a single balance run.
39 * Limited because this is done with IRQs disabled.
41 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
44 * period over which we average the RT time consumption, measured
49 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
52 * period over which we measure -rt task CPU usage in us.
55 unsigned int sysctl_sched_rt_period
= 1000000;
57 __read_mostly
int scheduler_running
;
60 * part of the period that we allow rt tasks to run in us.
63 int sysctl_sched_rt_runtime
= 950000;
66 * __task_rq_lock - lock the rq @p resides on.
68 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
73 lockdep_assert_held(&p
->pi_lock
);
77 raw_spin_lock(&rq
->lock
);
78 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
82 raw_spin_unlock(&rq
->lock
);
84 while (unlikely(task_on_rq_migrating(p
)))
90 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
92 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
93 __acquires(p
->pi_lock
)
99 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
101 raw_spin_lock(&rq
->lock
);
103 * move_queued_task() task_rq_lock()
106 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
107 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
108 * [S] ->cpu = new_cpu [L] task_rq()
112 * If we observe the old CPU in task_rq_lock, the acquire of
113 * the old rq->lock will fully serialize against the stores.
115 * If we observe the new CPU in task_rq_lock, the acquire will
116 * pair with the WMB to ensure we must then also see migrating.
118 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
122 raw_spin_unlock(&rq
->lock
);
123 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
125 while (unlikely(task_on_rq_migrating(p
)))
131 * RQ-clock updating methods:
134 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
137 * In theory, the compile should just see 0 here, and optimize out the call
138 * to sched_rt_avg_update. But I don't trust it...
140 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
141 s64 steal
= 0, irq_delta
= 0;
143 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
144 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
147 * Since irq_time is only updated on {soft,}irq_exit, we might run into
148 * this case when a previous update_rq_clock() happened inside a
151 * When this happens, we stop ->clock_task and only update the
152 * prev_irq_time stamp to account for the part that fit, so that a next
153 * update will consume the rest. This ensures ->clock_task is
156 * It does however cause some slight miss-attribution of {soft,}irq
157 * time, a more accurate solution would be to update the irq_time using
158 * the current rq->clock timestamp, except that would require using
161 if (irq_delta
> delta
)
164 rq
->prev_irq_time
+= irq_delta
;
167 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
168 if (static_key_false((¶virt_steal_rq_enabled
))) {
169 steal
= paravirt_steal_clock(cpu_of(rq
));
170 steal
-= rq
->prev_steal_time_rq
;
172 if (unlikely(steal
> delta
))
175 rq
->prev_steal_time_rq
+= steal
;
180 rq
->clock_task
+= delta
;
182 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
183 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
184 sched_rt_avg_update(rq
, irq_delta
+ steal
);
188 void update_rq_clock(struct rq
*rq
)
192 lockdep_assert_held(&rq
->lock
);
194 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
197 #ifdef CONFIG_SCHED_DEBUG
198 if (sched_feat(WARN_DOUBLE_CLOCK
))
199 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
200 rq
->clock_update_flags
|= RQCF_UPDATED
;
203 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
207 update_rq_clock_task(rq
, delta
);
211 #ifdef CONFIG_SCHED_HRTICK
213 * Use HR-timers to deliver accurate preemption points.
216 static void hrtick_clear(struct rq
*rq
)
218 if (hrtimer_active(&rq
->hrtick_timer
))
219 hrtimer_cancel(&rq
->hrtick_timer
);
223 * High-resolution timer tick.
224 * Runs from hardirq context with interrupts disabled.
226 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
228 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
231 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
235 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
238 return HRTIMER_NORESTART
;
243 static void __hrtick_restart(struct rq
*rq
)
245 struct hrtimer
*timer
= &rq
->hrtick_timer
;
247 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
251 * called from hardirq (IPI) context
253 static void __hrtick_start(void *arg
)
259 __hrtick_restart(rq
);
260 rq
->hrtick_csd_pending
= 0;
265 * Called to set the hrtick timer state.
267 * called with rq->lock held and irqs disabled
269 void hrtick_start(struct rq
*rq
, u64 delay
)
271 struct hrtimer
*timer
= &rq
->hrtick_timer
;
276 * Don't schedule slices shorter than 10000ns, that just
277 * doesn't make sense and can cause timer DoS.
279 delta
= max_t(s64
, delay
, 10000LL);
280 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
282 hrtimer_set_expires(timer
, time
);
284 if (rq
== this_rq()) {
285 __hrtick_restart(rq
);
286 } else if (!rq
->hrtick_csd_pending
) {
287 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
288 rq
->hrtick_csd_pending
= 1;
294 * Called to set the hrtick timer state.
296 * called with rq->lock held and irqs disabled
298 void hrtick_start(struct rq
*rq
, u64 delay
)
301 * Don't schedule slices shorter than 10000ns, that just
302 * doesn't make sense. Rely on vruntime for fairness.
304 delay
= max_t(u64
, delay
, 10000LL);
305 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
306 HRTIMER_MODE_REL_PINNED
);
308 #endif /* CONFIG_SMP */
310 static void hrtick_rq_init(struct rq
*rq
)
313 rq
->hrtick_csd_pending
= 0;
315 rq
->hrtick_csd
.flags
= 0;
316 rq
->hrtick_csd
.func
= __hrtick_start
;
317 rq
->hrtick_csd
.info
= rq
;
320 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
321 rq
->hrtick_timer
.function
= hrtick
;
323 #else /* CONFIG_SCHED_HRTICK */
324 static inline void hrtick_clear(struct rq
*rq
)
328 static inline void hrtick_rq_init(struct rq
*rq
)
331 #endif /* CONFIG_SCHED_HRTICK */
334 * cmpxchg based fetch_or, macro so it works for different integer types
336 #define fetch_or(ptr, mask) \
338 typeof(ptr) _ptr = (ptr); \
339 typeof(mask) _mask = (mask); \
340 typeof(*_ptr) _old, _val = *_ptr; \
343 _old = cmpxchg(_ptr, _val, _val | _mask); \
351 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
353 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
354 * this avoids any races wrt polling state changes and thereby avoids
357 static bool set_nr_and_not_polling(struct task_struct
*p
)
359 struct thread_info
*ti
= task_thread_info(p
);
360 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
364 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
366 * If this returns true, then the idle task promises to call
367 * sched_ttwu_pending() and reschedule soon.
369 static bool set_nr_if_polling(struct task_struct
*p
)
371 struct thread_info
*ti
= task_thread_info(p
);
372 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
375 if (!(val
& _TIF_POLLING_NRFLAG
))
377 if (val
& _TIF_NEED_RESCHED
)
379 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
388 static bool set_nr_and_not_polling(struct task_struct
*p
)
390 set_tsk_need_resched(p
);
395 static bool set_nr_if_polling(struct task_struct
*p
)
402 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
404 struct wake_q_node
*node
= &task
->wake_q
;
407 * Atomically grab the task, if ->wake_q is !nil already it means
408 * its already queued (either by us or someone else) and will get the
409 * wakeup due to that.
411 * This cmpxchg() implies a full barrier, which pairs with the write
412 * barrier implied by the wakeup in wake_up_q().
414 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
417 get_task_struct(task
);
420 * The head is context local, there can be no concurrency.
423 head
->lastp
= &node
->next
;
426 void wake_up_q(struct wake_q_head
*head
)
428 struct wake_q_node
*node
= head
->first
;
430 while (node
!= WAKE_Q_TAIL
) {
431 struct task_struct
*task
;
433 task
= container_of(node
, struct task_struct
, wake_q
);
435 /* Task can safely be re-inserted now: */
437 task
->wake_q
.next
= NULL
;
440 * wake_up_process() implies a wmb() to pair with the queueing
441 * in wake_q_add() so as not to miss wakeups.
443 wake_up_process(task
);
444 put_task_struct(task
);
449 * resched_curr - mark rq's current task 'to be rescheduled now'.
451 * On UP this means the setting of the need_resched flag, on SMP it
452 * might also involve a cross-CPU call to trigger the scheduler on
455 void resched_curr(struct rq
*rq
)
457 struct task_struct
*curr
= rq
->curr
;
460 lockdep_assert_held(&rq
->lock
);
462 if (test_tsk_need_resched(curr
))
467 if (cpu
== smp_processor_id()) {
468 set_tsk_need_resched(curr
);
469 set_preempt_need_resched();
473 if (set_nr_and_not_polling(curr
))
474 smp_send_reschedule(cpu
);
476 trace_sched_wake_idle_without_ipi(cpu
);
479 void resched_cpu(int cpu
)
481 struct rq
*rq
= cpu_rq(cpu
);
484 raw_spin_lock_irqsave(&rq
->lock
, flags
);
485 if (cpu_online(cpu
) || cpu
== smp_processor_id())
487 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
491 #ifdef CONFIG_NO_HZ_COMMON
493 * In the semi idle case, use the nearest busy CPU for migrating timers
494 * from an idle CPU. This is good for power-savings.
496 * We don't do similar optimization for completely idle system, as
497 * selecting an idle CPU will add more delays to the timers than intended
498 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
500 int get_nohz_timer_target(void)
502 int i
, cpu
= smp_processor_id();
503 struct sched_domain
*sd
;
505 if (!idle_cpu(cpu
) && housekeeping_cpu(cpu
, HK_FLAG_TIMER
))
509 for_each_domain(cpu
, sd
) {
510 for_each_cpu(i
, sched_domain_span(sd
)) {
514 if (!idle_cpu(i
) && housekeeping_cpu(i
, HK_FLAG_TIMER
)) {
521 if (!housekeeping_cpu(cpu
, HK_FLAG_TIMER
))
522 cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
529 * When add_timer_on() enqueues a timer into the timer wheel of an
530 * idle CPU then this timer might expire before the next timer event
531 * which is scheduled to wake up that CPU. In case of a completely
532 * idle system the next event might even be infinite time into the
533 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
534 * leaves the inner idle loop so the newly added timer is taken into
535 * account when the CPU goes back to idle and evaluates the timer
536 * wheel for the next timer event.
538 static void wake_up_idle_cpu(int cpu
)
540 struct rq
*rq
= cpu_rq(cpu
);
542 if (cpu
== smp_processor_id())
545 if (set_nr_and_not_polling(rq
->idle
))
546 smp_send_reschedule(cpu
);
548 trace_sched_wake_idle_without_ipi(cpu
);
551 static bool wake_up_full_nohz_cpu(int cpu
)
554 * We just need the target to call irq_exit() and re-evaluate
555 * the next tick. The nohz full kick at least implies that.
556 * If needed we can still optimize that later with an
559 if (cpu_is_offline(cpu
))
560 return true; /* Don't try to wake offline CPUs. */
561 if (tick_nohz_full_cpu(cpu
)) {
562 if (cpu
!= smp_processor_id() ||
563 tick_nohz_tick_stopped())
564 tick_nohz_full_kick_cpu(cpu
);
572 * Wake up the specified CPU. If the CPU is going offline, it is the
573 * caller's responsibility to deal with the lost wakeup, for example,
574 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
576 void wake_up_nohz_cpu(int cpu
)
578 if (!wake_up_full_nohz_cpu(cpu
))
579 wake_up_idle_cpu(cpu
);
582 static inline bool got_nohz_idle_kick(void)
584 int cpu
= smp_processor_id();
586 if (!(atomic_read(nohz_flags(cpu
)) & NOHZ_KICK_MASK
))
589 if (idle_cpu(cpu
) && !need_resched())
593 * We can't run Idle Load Balance on this CPU for this time so we
594 * cancel it and clear NOHZ_BALANCE_KICK
596 atomic_andnot(NOHZ_KICK_MASK
, nohz_flags(cpu
));
600 #else /* CONFIG_NO_HZ_COMMON */
602 static inline bool got_nohz_idle_kick(void)
607 #endif /* CONFIG_NO_HZ_COMMON */
609 #ifdef CONFIG_NO_HZ_FULL
610 bool sched_can_stop_tick(struct rq
*rq
)
614 /* Deadline tasks, even if single, need the tick */
615 if (rq
->dl
.dl_nr_running
)
619 * If there are more than one RR tasks, we need the tick to effect the
620 * actual RR behaviour.
622 if (rq
->rt
.rr_nr_running
) {
623 if (rq
->rt
.rr_nr_running
== 1)
630 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
631 * forced preemption between FIFO tasks.
633 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
638 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
639 * if there's more than one we need the tick for involuntary
642 if (rq
->nr_running
> 1)
647 #endif /* CONFIG_NO_HZ_FULL */
649 void sched_avg_update(struct rq
*rq
)
651 s64 period
= sched_avg_period();
653 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
655 * Inline assembly required to prevent the compiler
656 * optimising this loop into a divmod call.
657 * See __iter_div_u64_rem() for another example of this.
659 asm("" : "+rm" (rq
->age_stamp
));
660 rq
->age_stamp
+= period
;
665 #endif /* CONFIG_SMP */
667 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
668 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
670 * Iterate task_group tree rooted at *from, calling @down when first entering a
671 * node and @up when leaving it for the final time.
673 * Caller must hold rcu_lock or sufficient equivalent.
675 int walk_tg_tree_from(struct task_group
*from
,
676 tg_visitor down
, tg_visitor up
, void *data
)
678 struct task_group
*parent
, *child
;
684 ret
= (*down
)(parent
, data
);
687 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
694 ret
= (*up
)(parent
, data
);
695 if (ret
|| parent
== from
)
699 parent
= parent
->parent
;
706 int tg_nop(struct task_group
*tg
, void *data
)
712 static void set_load_weight(struct task_struct
*p
, bool update_load
)
714 int prio
= p
->static_prio
- MAX_RT_PRIO
;
715 struct load_weight
*load
= &p
->se
.load
;
718 * SCHED_IDLE tasks get minimal weight:
720 if (idle_policy(p
->policy
)) {
721 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
722 load
->inv_weight
= WMULT_IDLEPRIO
;
727 * SCHED_OTHER tasks have to update their load when changing their
730 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
731 reweight_task(p
, prio
);
733 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
734 load
->inv_weight
= sched_prio_to_wmult
[prio
];
738 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
740 if (!(flags
& ENQUEUE_NOCLOCK
))
743 if (!(flags
& ENQUEUE_RESTORE
))
744 sched_info_queued(rq
, p
);
746 p
->sched_class
->enqueue_task(rq
, p
, flags
);
749 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
751 if (!(flags
& DEQUEUE_NOCLOCK
))
754 if (!(flags
& DEQUEUE_SAVE
))
755 sched_info_dequeued(rq
, p
);
757 p
->sched_class
->dequeue_task(rq
, p
, flags
);
760 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
762 if (task_contributes_to_load(p
))
763 rq
->nr_uninterruptible
--;
765 enqueue_task(rq
, p
, flags
);
768 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
770 if (task_contributes_to_load(p
))
771 rq
->nr_uninterruptible
++;
773 dequeue_task(rq
, p
, flags
);
777 * __normal_prio - return the priority that is based on the static prio
779 static inline int __normal_prio(struct task_struct
*p
)
781 return p
->static_prio
;
785 * Calculate the expected normal priority: i.e. priority
786 * without taking RT-inheritance into account. Might be
787 * boosted by interactivity modifiers. Changes upon fork,
788 * setprio syscalls, and whenever the interactivity
789 * estimator recalculates.
791 static inline int normal_prio(struct task_struct
*p
)
795 if (task_has_dl_policy(p
))
796 prio
= MAX_DL_PRIO
-1;
797 else if (task_has_rt_policy(p
))
798 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
800 prio
= __normal_prio(p
);
805 * Calculate the current priority, i.e. the priority
806 * taken into account by the scheduler. This value might
807 * be boosted by RT tasks, or might be boosted by
808 * interactivity modifiers. Will be RT if the task got
809 * RT-boosted. If not then it returns p->normal_prio.
811 static int effective_prio(struct task_struct
*p
)
813 p
->normal_prio
= normal_prio(p
);
815 * If we are RT tasks or we were boosted to RT priority,
816 * keep the priority unchanged. Otherwise, update priority
817 * to the normal priority:
819 if (!rt_prio(p
->prio
))
820 return p
->normal_prio
;
825 * task_curr - is this task currently executing on a CPU?
826 * @p: the task in question.
828 * Return: 1 if the task is currently executing. 0 otherwise.
830 inline int task_curr(const struct task_struct
*p
)
832 return cpu_curr(task_cpu(p
)) == p
;
836 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
837 * use the balance_callback list if you want balancing.
839 * this means any call to check_class_changed() must be followed by a call to
840 * balance_callback().
842 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
843 const struct sched_class
*prev_class
,
846 if (prev_class
!= p
->sched_class
) {
847 if (prev_class
->switched_from
)
848 prev_class
->switched_from(rq
, p
);
850 p
->sched_class
->switched_to(rq
, p
);
851 } else if (oldprio
!= p
->prio
|| dl_task(p
))
852 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
855 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
857 const struct sched_class
*class;
859 if (p
->sched_class
== rq
->curr
->sched_class
) {
860 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
862 for_each_class(class) {
863 if (class == rq
->curr
->sched_class
)
865 if (class == p
->sched_class
) {
873 * A queue event has occurred, and we're going to schedule. In
874 * this case, we can save a useless back to back clock update.
876 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
877 rq_clock_skip_update(rq
);
882 * This is how migration works:
884 * 1) we invoke migration_cpu_stop() on the target CPU using
886 * 2) stopper starts to run (implicitly forcing the migrated thread
888 * 3) it checks whether the migrated task is still in the wrong runqueue.
889 * 4) if it's in the wrong runqueue then the migration thread removes
890 * it and puts it into the right queue.
891 * 5) stopper completes and stop_one_cpu() returns and the migration
896 * move_queued_task - move a queued task to new rq.
898 * Returns (locked) new rq. Old rq's lock is released.
900 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
901 struct task_struct
*p
, int new_cpu
)
903 lockdep_assert_held(&rq
->lock
);
905 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
906 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
907 set_task_cpu(p
, new_cpu
);
910 rq
= cpu_rq(new_cpu
);
913 BUG_ON(task_cpu(p
) != new_cpu
);
914 enqueue_task(rq
, p
, 0);
915 p
->on_rq
= TASK_ON_RQ_QUEUED
;
916 check_preempt_curr(rq
, p
, 0);
921 struct migration_arg
{
922 struct task_struct
*task
;
927 * Move (not current) task off this CPU, onto the destination CPU. We're doing
928 * this because either it can't run here any more (set_cpus_allowed()
929 * away from this CPU, or CPU going down), or because we're
930 * attempting to rebalance this task on exec (sched_exec).
932 * So we race with normal scheduler movements, but that's OK, as long
933 * as the task is no longer on this CPU.
935 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
936 struct task_struct
*p
, int dest_cpu
)
938 if (p
->flags
& PF_KTHREAD
) {
939 if (unlikely(!cpu_online(dest_cpu
)))
942 if (unlikely(!cpu_active(dest_cpu
)))
946 /* Affinity changed (again). */
947 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
951 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
957 * migration_cpu_stop - this will be executed by a highprio stopper thread
958 * and performs thread migration by bumping thread off CPU then
959 * 'pushing' onto another runqueue.
961 static int migration_cpu_stop(void *data
)
963 struct migration_arg
*arg
= data
;
964 struct task_struct
*p
= arg
->task
;
965 struct rq
*rq
= this_rq();
969 * The original target CPU might have gone down and we might
970 * be on another CPU but it doesn't matter.
974 * We need to explicitly wake pending tasks before running
975 * __migrate_task() such that we will not miss enforcing cpus_allowed
976 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
978 sched_ttwu_pending();
980 raw_spin_lock(&p
->pi_lock
);
983 * If task_rq(p) != rq, it cannot be migrated here, because we're
984 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
985 * we're holding p->pi_lock.
987 if (task_rq(p
) == rq
) {
988 if (task_on_rq_queued(p
))
989 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
991 p
->wake_cpu
= arg
->dest_cpu
;
994 raw_spin_unlock(&p
->pi_lock
);
1001 * sched_class::set_cpus_allowed must do the below, but is not required to
1002 * actually call this function.
1004 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1006 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1007 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1010 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1012 struct rq
*rq
= task_rq(p
);
1013 bool queued
, running
;
1015 lockdep_assert_held(&p
->pi_lock
);
1017 queued
= task_on_rq_queued(p
);
1018 running
= task_current(rq
, p
);
1022 * Because __kthread_bind() calls this on blocked tasks without
1025 lockdep_assert_held(&rq
->lock
);
1026 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1029 put_prev_task(rq
, p
);
1031 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1034 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1036 set_curr_task(rq
, p
);
1040 * Change a given task's CPU affinity. Migrate the thread to a
1041 * proper CPU and schedule it away if the CPU it's executing on
1042 * is removed from the allowed bitmask.
1044 * NOTE: the caller must have a valid reference to the task, the
1045 * task must not exit() & deallocate itself prematurely. The
1046 * call is not atomic; no spinlocks may be held.
1048 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1049 const struct cpumask
*new_mask
, bool check
)
1051 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1052 unsigned int dest_cpu
;
1057 rq
= task_rq_lock(p
, &rf
);
1058 update_rq_clock(rq
);
1060 if (p
->flags
& PF_KTHREAD
) {
1062 * Kernel threads are allowed on online && !active CPUs
1064 cpu_valid_mask
= cpu_online_mask
;
1068 * Must re-check here, to close a race against __kthread_bind(),
1069 * sched_setaffinity() is not guaranteed to observe the flag.
1071 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1076 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1079 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1084 do_set_cpus_allowed(p
, new_mask
);
1086 if (p
->flags
& PF_KTHREAD
) {
1088 * For kernel threads that do indeed end up on online &&
1089 * !active we want to ensure they are strict per-CPU threads.
1091 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1092 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1093 p
->nr_cpus_allowed
!= 1);
1096 /* Can the task run on the task's current CPU? If so, we're done */
1097 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1100 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1101 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1102 struct migration_arg arg
= { p
, dest_cpu
};
1103 /* Need help from migration thread: drop lock and wait. */
1104 task_rq_unlock(rq
, p
, &rf
);
1105 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1106 tlb_migrate_finish(p
->mm
);
1108 } else if (task_on_rq_queued(p
)) {
1110 * OK, since we're going to drop the lock immediately
1111 * afterwards anyway.
1113 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1116 task_rq_unlock(rq
, p
, &rf
);
1121 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1123 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1125 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1127 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1129 #ifdef CONFIG_SCHED_DEBUG
1131 * We should never call set_task_cpu() on a blocked task,
1132 * ttwu() will sort out the placement.
1134 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1138 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1139 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1140 * time relying on p->on_rq.
1142 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1143 p
->sched_class
== &fair_sched_class
&&
1144 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1146 #ifdef CONFIG_LOCKDEP
1148 * The caller should hold either p->pi_lock or rq->lock, when changing
1149 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1151 * sched_move_task() holds both and thus holding either pins the cgroup,
1154 * Furthermore, all task_rq users should acquire both locks, see
1157 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1158 lockdep_is_held(&task_rq(p
)->lock
)));
1161 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1163 WARN_ON_ONCE(!cpu_online(new_cpu
));
1166 trace_sched_migrate_task(p
, new_cpu
);
1168 if (task_cpu(p
) != new_cpu
) {
1169 if (p
->sched_class
->migrate_task_rq
)
1170 p
->sched_class
->migrate_task_rq(p
);
1171 p
->se
.nr_migrations
++;
1172 perf_event_task_migrate(p
);
1175 __set_task_cpu(p
, new_cpu
);
1178 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1180 if (task_on_rq_queued(p
)) {
1181 struct rq
*src_rq
, *dst_rq
;
1182 struct rq_flags srf
, drf
;
1184 src_rq
= task_rq(p
);
1185 dst_rq
= cpu_rq(cpu
);
1187 rq_pin_lock(src_rq
, &srf
);
1188 rq_pin_lock(dst_rq
, &drf
);
1190 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1191 deactivate_task(src_rq
, p
, 0);
1192 set_task_cpu(p
, cpu
);
1193 activate_task(dst_rq
, p
, 0);
1194 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1195 check_preempt_curr(dst_rq
, p
, 0);
1197 rq_unpin_lock(dst_rq
, &drf
);
1198 rq_unpin_lock(src_rq
, &srf
);
1202 * Task isn't running anymore; make it appear like we migrated
1203 * it before it went to sleep. This means on wakeup we make the
1204 * previous CPU our target instead of where it really is.
1210 struct migration_swap_arg
{
1211 struct task_struct
*src_task
, *dst_task
;
1212 int src_cpu
, dst_cpu
;
1215 static int migrate_swap_stop(void *data
)
1217 struct migration_swap_arg
*arg
= data
;
1218 struct rq
*src_rq
, *dst_rq
;
1221 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1224 src_rq
= cpu_rq(arg
->src_cpu
);
1225 dst_rq
= cpu_rq(arg
->dst_cpu
);
1227 double_raw_lock(&arg
->src_task
->pi_lock
,
1228 &arg
->dst_task
->pi_lock
);
1229 double_rq_lock(src_rq
, dst_rq
);
1231 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1234 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1237 if (!cpumask_test_cpu(arg
->dst_cpu
, &arg
->src_task
->cpus_allowed
))
1240 if (!cpumask_test_cpu(arg
->src_cpu
, &arg
->dst_task
->cpus_allowed
))
1243 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1244 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1249 double_rq_unlock(src_rq
, dst_rq
);
1250 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1251 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1257 * Cross migrate two tasks
1259 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1261 struct migration_swap_arg arg
;
1264 arg
= (struct migration_swap_arg
){
1266 .src_cpu
= task_cpu(cur
),
1268 .dst_cpu
= task_cpu(p
),
1271 if (arg
.src_cpu
== arg
.dst_cpu
)
1275 * These three tests are all lockless; this is OK since all of them
1276 * will be re-checked with proper locks held further down the line.
1278 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1281 if (!cpumask_test_cpu(arg
.dst_cpu
, &arg
.src_task
->cpus_allowed
))
1284 if (!cpumask_test_cpu(arg
.src_cpu
, &arg
.dst_task
->cpus_allowed
))
1287 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1288 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1295 * wait_task_inactive - wait for a thread to unschedule.
1297 * If @match_state is nonzero, it's the @p->state value just checked and
1298 * not expected to change. If it changes, i.e. @p might have woken up,
1299 * then return zero. When we succeed in waiting for @p to be off its CPU,
1300 * we return a positive number (its total switch count). If a second call
1301 * a short while later returns the same number, the caller can be sure that
1302 * @p has remained unscheduled the whole time.
1304 * The caller must ensure that the task *will* unschedule sometime soon,
1305 * else this function might spin for a *long* time. This function can't
1306 * be called with interrupts off, or it may introduce deadlock with
1307 * smp_call_function() if an IPI is sent by the same process we are
1308 * waiting to become inactive.
1310 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1312 int running
, queued
;
1319 * We do the initial early heuristics without holding
1320 * any task-queue locks at all. We'll only try to get
1321 * the runqueue lock when things look like they will
1327 * If the task is actively running on another CPU
1328 * still, just relax and busy-wait without holding
1331 * NOTE! Since we don't hold any locks, it's not
1332 * even sure that "rq" stays as the right runqueue!
1333 * But we don't care, since "task_running()" will
1334 * return false if the runqueue has changed and p
1335 * is actually now running somewhere else!
1337 while (task_running(rq
, p
)) {
1338 if (match_state
&& unlikely(p
->state
!= match_state
))
1344 * Ok, time to look more closely! We need the rq
1345 * lock now, to be *sure*. If we're wrong, we'll
1346 * just go back and repeat.
1348 rq
= task_rq_lock(p
, &rf
);
1349 trace_sched_wait_task(p
);
1350 running
= task_running(rq
, p
);
1351 queued
= task_on_rq_queued(p
);
1353 if (!match_state
|| p
->state
== match_state
)
1354 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1355 task_rq_unlock(rq
, p
, &rf
);
1358 * If it changed from the expected state, bail out now.
1360 if (unlikely(!ncsw
))
1364 * Was it really running after all now that we
1365 * checked with the proper locks actually held?
1367 * Oops. Go back and try again..
1369 if (unlikely(running
)) {
1375 * It's not enough that it's not actively running,
1376 * it must be off the runqueue _entirely_, and not
1379 * So if it was still runnable (but just not actively
1380 * running right now), it's preempted, and we should
1381 * yield - it could be a while.
1383 if (unlikely(queued
)) {
1384 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1386 set_current_state(TASK_UNINTERRUPTIBLE
);
1387 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1392 * Ahh, all good. It wasn't running, and it wasn't
1393 * runnable, which means that it will never become
1394 * running in the future either. We're all done!
1403 * kick_process - kick a running thread to enter/exit the kernel
1404 * @p: the to-be-kicked thread
1406 * Cause a process which is running on another CPU to enter
1407 * kernel-mode, without any delay. (to get signals handled.)
1409 * NOTE: this function doesn't have to take the runqueue lock,
1410 * because all it wants to ensure is that the remote task enters
1411 * the kernel. If the IPI races and the task has been migrated
1412 * to another CPU then no harm is done and the purpose has been
1415 void kick_process(struct task_struct
*p
)
1421 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1422 smp_send_reschedule(cpu
);
1425 EXPORT_SYMBOL_GPL(kick_process
);
1428 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1430 * A few notes on cpu_active vs cpu_online:
1432 * - cpu_active must be a subset of cpu_online
1434 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1435 * see __set_cpus_allowed_ptr(). At this point the newly online
1436 * CPU isn't yet part of the sched domains, and balancing will not
1439 * - on CPU-down we clear cpu_active() to mask the sched domains and
1440 * avoid the load balancer to place new tasks on the to be removed
1441 * CPU. Existing tasks will remain running there and will be taken
1444 * This means that fallback selection must not select !active CPUs.
1445 * And can assume that any active CPU must be online. Conversely
1446 * select_task_rq() below may allow selection of !active CPUs in order
1447 * to satisfy the above rules.
1449 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1451 int nid
= cpu_to_node(cpu
);
1452 const struct cpumask
*nodemask
= NULL
;
1453 enum { cpuset
, possible
, fail
} state
= cpuset
;
1457 * If the node that the CPU is on has been offlined, cpu_to_node()
1458 * will return -1. There is no CPU on the node, and we should
1459 * select the CPU on the other node.
1462 nodemask
= cpumask_of_node(nid
);
1464 /* Look for allowed, online CPU in same node. */
1465 for_each_cpu(dest_cpu
, nodemask
) {
1466 if (!cpu_active(dest_cpu
))
1468 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
1474 /* Any allowed, online CPU? */
1475 for_each_cpu(dest_cpu
, &p
->cpus_allowed
) {
1476 if (!(p
->flags
& PF_KTHREAD
) && !cpu_active(dest_cpu
))
1478 if (!cpu_online(dest_cpu
))
1483 /* No more Mr. Nice Guy. */
1486 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1487 cpuset_cpus_allowed_fallback(p
);
1493 do_set_cpus_allowed(p
, cpu_possible_mask
);
1504 if (state
!= cpuset
) {
1506 * Don't tell them about moving exiting tasks or
1507 * kernel threads (both mm NULL), since they never
1510 if (p
->mm
&& printk_ratelimit()) {
1511 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1512 task_pid_nr(p
), p
->comm
, cpu
);
1520 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1523 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1525 lockdep_assert_held(&p
->pi_lock
);
1527 if (p
->nr_cpus_allowed
> 1)
1528 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1530 cpu
= cpumask_any(&p
->cpus_allowed
);
1533 * In order not to call set_task_cpu() on a blocking task we need
1534 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1537 * Since this is common to all placement strategies, this lives here.
1539 * [ this allows ->select_task() to simply return task_cpu(p) and
1540 * not worry about this generic constraint ]
1542 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
1544 cpu
= select_fallback_rq(task_cpu(p
), p
);
1549 static void update_avg(u64
*avg
, u64 sample
)
1551 s64 diff
= sample
- *avg
;
1555 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
1557 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
1558 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
1562 * Make it appear like a SCHED_FIFO task, its something
1563 * userspace knows about and won't get confused about.
1565 * Also, it will make PI more or less work without too
1566 * much confusion -- but then, stop work should not
1567 * rely on PI working anyway.
1569 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
1571 stop
->sched_class
= &stop_sched_class
;
1574 cpu_rq(cpu
)->stop
= stop
;
1578 * Reset it back to a normal scheduling class so that
1579 * it can die in pieces.
1581 old_stop
->sched_class
= &rt_sched_class
;
1587 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1588 const struct cpumask
*new_mask
, bool check
)
1590 return set_cpus_allowed_ptr(p
, new_mask
);
1593 #endif /* CONFIG_SMP */
1596 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1600 if (!schedstat_enabled())
1606 if (cpu
== rq
->cpu
) {
1607 __schedstat_inc(rq
->ttwu_local
);
1608 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
1610 struct sched_domain
*sd
;
1612 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
1614 for_each_domain(rq
->cpu
, sd
) {
1615 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1616 __schedstat_inc(sd
->ttwu_wake_remote
);
1623 if (wake_flags
& WF_MIGRATED
)
1624 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
1625 #endif /* CONFIG_SMP */
1627 __schedstat_inc(rq
->ttwu_count
);
1628 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
1630 if (wake_flags
& WF_SYNC
)
1631 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
1634 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1636 activate_task(rq
, p
, en_flags
);
1637 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1639 /* If a worker is waking up, notify the workqueue: */
1640 if (p
->flags
& PF_WQ_WORKER
)
1641 wq_worker_waking_up(p
, cpu_of(rq
));
1645 * Mark the task runnable and perform wakeup-preemption.
1647 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1648 struct rq_flags
*rf
)
1650 check_preempt_curr(rq
, p
, wake_flags
);
1651 p
->state
= TASK_RUNNING
;
1652 trace_sched_wakeup(p
);
1655 if (p
->sched_class
->task_woken
) {
1657 * Our task @p is fully woken up and running; so its safe to
1658 * drop the rq->lock, hereafter rq is only used for statistics.
1660 rq_unpin_lock(rq
, rf
);
1661 p
->sched_class
->task_woken(rq
, p
);
1662 rq_repin_lock(rq
, rf
);
1665 if (rq
->idle_stamp
) {
1666 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1667 u64 max
= 2*rq
->max_idle_balance_cost
;
1669 update_avg(&rq
->avg_idle
, delta
);
1671 if (rq
->avg_idle
> max
)
1680 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1681 struct rq_flags
*rf
)
1683 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
1685 lockdep_assert_held(&rq
->lock
);
1688 if (p
->sched_contributes_to_load
)
1689 rq
->nr_uninterruptible
--;
1691 if (wake_flags
& WF_MIGRATED
)
1692 en_flags
|= ENQUEUE_MIGRATED
;
1695 ttwu_activate(rq
, p
, en_flags
);
1696 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
1700 * Called in case the task @p isn't fully descheduled from its runqueue,
1701 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1702 * since all we need to do is flip p->state to TASK_RUNNING, since
1703 * the task is still ->on_rq.
1705 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1711 rq
= __task_rq_lock(p
, &rf
);
1712 if (task_on_rq_queued(p
)) {
1713 /* check_preempt_curr() may use rq clock */
1714 update_rq_clock(rq
);
1715 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
1718 __task_rq_unlock(rq
, &rf
);
1724 void sched_ttwu_pending(void)
1726 struct rq
*rq
= this_rq();
1727 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1728 struct task_struct
*p
, *t
;
1734 rq_lock_irqsave(rq
, &rf
);
1735 update_rq_clock(rq
);
1737 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
)
1738 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
1740 rq_unlock_irqrestore(rq
, &rf
);
1743 void scheduler_ipi(void)
1746 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1747 * TIF_NEED_RESCHED remotely (for the first time) will also send
1750 preempt_fold_need_resched();
1752 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1756 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1757 * traditionally all their work was done from the interrupt return
1758 * path. Now that we actually do some work, we need to make sure
1761 * Some archs already do call them, luckily irq_enter/exit nest
1764 * Arguably we should visit all archs and update all handlers,
1765 * however a fair share of IPIs are still resched only so this would
1766 * somewhat pessimize the simple resched case.
1769 sched_ttwu_pending();
1772 * Check if someone kicked us for doing the nohz idle load balance.
1774 if (unlikely(got_nohz_idle_kick())) {
1775 this_rq()->idle_balance
= 1;
1776 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1781 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1783 struct rq
*rq
= cpu_rq(cpu
);
1785 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1787 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1788 if (!set_nr_if_polling(rq
->idle
))
1789 smp_send_reschedule(cpu
);
1791 trace_sched_wake_idle_without_ipi(cpu
);
1795 void wake_up_if_idle(int cpu
)
1797 struct rq
*rq
= cpu_rq(cpu
);
1802 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1805 if (set_nr_if_polling(rq
->idle
)) {
1806 trace_sched_wake_idle_without_ipi(cpu
);
1808 rq_lock_irqsave(rq
, &rf
);
1809 if (is_idle_task(rq
->curr
))
1810 smp_send_reschedule(cpu
);
1811 /* Else CPU is not idle, do nothing here: */
1812 rq_unlock_irqrestore(rq
, &rf
);
1819 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1821 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1823 #endif /* CONFIG_SMP */
1825 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1827 struct rq
*rq
= cpu_rq(cpu
);
1830 #if defined(CONFIG_SMP)
1831 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1832 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
1833 ttwu_queue_remote(p
, cpu
, wake_flags
);
1839 update_rq_clock(rq
);
1840 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
1845 * Notes on Program-Order guarantees on SMP systems.
1849 * The basic program-order guarantee on SMP systems is that when a task [t]
1850 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1851 * execution on its new CPU [c1].
1853 * For migration (of runnable tasks) this is provided by the following means:
1855 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1856 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1857 * rq(c1)->lock (if not at the same time, then in that order).
1858 * C) LOCK of the rq(c1)->lock scheduling in task
1860 * Transitivity guarantees that B happens after A and C after B.
1861 * Note: we only require RCpc transitivity.
1862 * Note: the CPU doing B need not be c0 or c1
1871 * UNLOCK rq(0)->lock
1873 * LOCK rq(0)->lock // orders against CPU0
1875 * UNLOCK rq(0)->lock
1879 * UNLOCK rq(1)->lock
1881 * LOCK rq(1)->lock // orders against CPU2
1884 * UNLOCK rq(1)->lock
1887 * BLOCKING -- aka. SLEEP + WAKEUP
1889 * For blocking we (obviously) need to provide the same guarantee as for
1890 * migration. However the means are completely different as there is no lock
1891 * chain to provide order. Instead we do:
1893 * 1) smp_store_release(X->on_cpu, 0)
1894 * 2) smp_cond_load_acquire(!X->on_cpu)
1898 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1900 * LOCK rq(0)->lock LOCK X->pi_lock
1903 * smp_store_release(X->on_cpu, 0);
1905 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1911 * X->state = RUNNING
1912 * UNLOCK rq(2)->lock
1914 * LOCK rq(2)->lock // orders against CPU1
1917 * UNLOCK rq(2)->lock
1920 * UNLOCK rq(0)->lock
1923 * However; for wakeups there is a second guarantee we must provide, namely we
1924 * must observe the state that lead to our wakeup. That is, not only must our
1925 * task observe its own prior state, it must also observe the stores prior to
1928 * This means that any means of doing remote wakeups must order the CPU doing
1929 * the wakeup against the CPU the task is going to end up running on. This,
1930 * however, is already required for the regular Program-Order guarantee above,
1931 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1936 * try_to_wake_up - wake up a thread
1937 * @p: the thread to be awakened
1938 * @state: the mask of task states that can be woken
1939 * @wake_flags: wake modifier flags (WF_*)
1941 * If (@state & @p->state) @p->state = TASK_RUNNING.
1943 * If the task was not queued/runnable, also place it back on a runqueue.
1945 * Atomic against schedule() which would dequeue a task, also see
1946 * set_current_state().
1948 * Return: %true if @p->state changes (an actual wakeup was done),
1952 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1954 unsigned long flags
;
1955 int cpu
, success
= 0;
1958 * If we are going to wake up a thread waiting for CONDITION we
1959 * need to ensure that CONDITION=1 done by the caller can not be
1960 * reordered with p->state check below. This pairs with mb() in
1961 * set_current_state() the waiting thread does.
1963 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1964 smp_mb__after_spinlock();
1965 if (!(p
->state
& state
))
1968 trace_sched_waking(p
);
1970 /* We're going to change ->state: */
1975 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1976 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1977 * in smp_cond_load_acquire() below.
1979 * sched_ttwu_pending() try_to_wake_up()
1980 * [S] p->on_rq = 1; [L] P->state
1981 * UNLOCK rq->lock -----.
1985 * LOCK rq->lock -----'
1989 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1991 * Pairs with the UNLOCK+LOCK on rq->lock from the
1992 * last wakeup of our task and the schedule that got our task
1996 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2001 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2002 * possible to, falsely, observe p->on_cpu == 0.
2004 * One must be running (->on_cpu == 1) in order to remove oneself
2005 * from the runqueue.
2007 * [S] ->on_cpu = 1; [L] ->on_rq
2011 * [S] ->on_rq = 0; [L] ->on_cpu
2013 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2014 * from the consecutive calls to schedule(); the first switching to our
2015 * task, the second putting it to sleep.
2020 * If the owning (remote) CPU is still in the middle of schedule() with
2021 * this task as prev, wait until its done referencing the task.
2023 * Pairs with the smp_store_release() in finish_task().
2025 * This ensures that tasks getting woken will be fully ordered against
2026 * their previous state and preserve Program Order.
2028 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2030 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2031 p
->state
= TASK_WAKING
;
2034 delayacct_blkio_end(p
);
2035 atomic_dec(&task_rq(p
)->nr_iowait
);
2038 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2039 if (task_cpu(p
) != cpu
) {
2040 wake_flags
|= WF_MIGRATED
;
2041 set_task_cpu(p
, cpu
);
2044 #else /* CONFIG_SMP */
2047 delayacct_blkio_end(p
);
2048 atomic_dec(&task_rq(p
)->nr_iowait
);
2051 #endif /* CONFIG_SMP */
2053 ttwu_queue(p
, cpu
, wake_flags
);
2055 ttwu_stat(p
, cpu
, wake_flags
);
2057 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2063 * try_to_wake_up_local - try to wake up a local task with rq lock held
2064 * @p: the thread to be awakened
2065 * @rf: request-queue flags for pinning
2067 * Put @p on the run-queue if it's not already there. The caller must
2068 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2071 static void try_to_wake_up_local(struct task_struct
*p
, struct rq_flags
*rf
)
2073 struct rq
*rq
= task_rq(p
);
2075 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2076 WARN_ON_ONCE(p
== current
))
2079 lockdep_assert_held(&rq
->lock
);
2081 if (!raw_spin_trylock(&p
->pi_lock
)) {
2083 * This is OK, because current is on_cpu, which avoids it being
2084 * picked for load-balance and preemption/IRQs are still
2085 * disabled avoiding further scheduler activity on it and we've
2086 * not yet picked a replacement task.
2089 raw_spin_lock(&p
->pi_lock
);
2093 if (!(p
->state
& TASK_NORMAL
))
2096 trace_sched_waking(p
);
2098 if (!task_on_rq_queued(p
)) {
2100 delayacct_blkio_end(p
);
2101 atomic_dec(&rq
->nr_iowait
);
2103 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
);
2106 ttwu_do_wakeup(rq
, p
, 0, rf
);
2107 ttwu_stat(p
, smp_processor_id(), 0);
2109 raw_spin_unlock(&p
->pi_lock
);
2113 * wake_up_process - Wake up a specific process
2114 * @p: The process to be woken up.
2116 * Attempt to wake up the nominated process and move it to the set of runnable
2119 * Return: 1 if the process was woken up, 0 if it was already running.
2121 * It may be assumed that this function implies a write memory barrier before
2122 * changing the task state if and only if any tasks are woken up.
2124 int wake_up_process(struct task_struct
*p
)
2126 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2128 EXPORT_SYMBOL(wake_up_process
);
2130 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2132 return try_to_wake_up(p
, state
, 0);
2136 * Perform scheduler related setup for a newly forked process p.
2137 * p is forked by current.
2139 * __sched_fork() is basic setup used by init_idle() too:
2141 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2146 p
->se
.exec_start
= 0;
2147 p
->se
.sum_exec_runtime
= 0;
2148 p
->se
.prev_sum_exec_runtime
= 0;
2149 p
->se
.nr_migrations
= 0;
2151 INIT_LIST_HEAD(&p
->se
.group_node
);
2153 #ifdef CONFIG_FAIR_GROUP_SCHED
2154 p
->se
.cfs_rq
= NULL
;
2157 #ifdef CONFIG_SCHEDSTATS
2158 /* Even if schedstat is disabled, there should not be garbage */
2159 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2162 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2163 init_dl_task_timer(&p
->dl
);
2164 init_dl_inactive_task_timer(&p
->dl
);
2165 __dl_clear_params(p
);
2167 INIT_LIST_HEAD(&p
->rt
.run_list
);
2169 p
->rt
.time_slice
= sched_rr_timeslice
;
2173 #ifdef CONFIG_PREEMPT_NOTIFIERS
2174 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2177 #ifdef CONFIG_NUMA_BALANCING
2178 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2179 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2180 p
->mm
->numa_scan_seq
= 0;
2183 if (clone_flags
& CLONE_VM
)
2184 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2186 p
->numa_preferred_nid
= -1;
2188 p
->node_stamp
= 0ULL;
2189 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2190 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2191 p
->numa_work
.next
= &p
->numa_work
;
2192 p
->numa_faults
= NULL
;
2193 p
->last_task_numa_placement
= 0;
2194 p
->last_sum_exec_runtime
= 0;
2196 p
->numa_group
= NULL
;
2197 #endif /* CONFIG_NUMA_BALANCING */
2200 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2202 #ifdef CONFIG_NUMA_BALANCING
2204 void set_numabalancing_state(bool enabled
)
2207 static_branch_enable(&sched_numa_balancing
);
2209 static_branch_disable(&sched_numa_balancing
);
2212 #ifdef CONFIG_PROC_SYSCTL
2213 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2214 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2218 int state
= static_branch_likely(&sched_numa_balancing
);
2220 if (write
&& !capable(CAP_SYS_ADMIN
))
2225 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2229 set_numabalancing_state(state
);
2235 #ifdef CONFIG_SCHEDSTATS
2237 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2238 static bool __initdata __sched_schedstats
= false;
2240 static void set_schedstats(bool enabled
)
2243 static_branch_enable(&sched_schedstats
);
2245 static_branch_disable(&sched_schedstats
);
2248 void force_schedstat_enabled(void)
2250 if (!schedstat_enabled()) {
2251 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2252 static_branch_enable(&sched_schedstats
);
2256 static int __init
setup_schedstats(char *str
)
2263 * This code is called before jump labels have been set up, so we can't
2264 * change the static branch directly just yet. Instead set a temporary
2265 * variable so init_schedstats() can do it later.
2267 if (!strcmp(str
, "enable")) {
2268 __sched_schedstats
= true;
2270 } else if (!strcmp(str
, "disable")) {
2271 __sched_schedstats
= false;
2276 pr_warn("Unable to parse schedstats=\n");
2280 __setup("schedstats=", setup_schedstats
);
2282 static void __init
init_schedstats(void)
2284 set_schedstats(__sched_schedstats
);
2287 #ifdef CONFIG_PROC_SYSCTL
2288 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2289 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2293 int state
= static_branch_likely(&sched_schedstats
);
2295 if (write
&& !capable(CAP_SYS_ADMIN
))
2300 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2304 set_schedstats(state
);
2307 #endif /* CONFIG_PROC_SYSCTL */
2308 #else /* !CONFIG_SCHEDSTATS */
2309 static inline void init_schedstats(void) {}
2310 #endif /* CONFIG_SCHEDSTATS */
2313 * fork()/clone()-time setup:
2315 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2317 unsigned long flags
;
2318 int cpu
= get_cpu();
2320 __sched_fork(clone_flags
, p
);
2322 * We mark the process as NEW here. This guarantees that
2323 * nobody will actually run it, and a signal or other external
2324 * event cannot wake it up and insert it on the runqueue either.
2326 p
->state
= TASK_NEW
;
2329 * Make sure we do not leak PI boosting priority to the child.
2331 p
->prio
= current
->normal_prio
;
2334 * Revert to default priority/policy on fork if requested.
2336 if (unlikely(p
->sched_reset_on_fork
)) {
2337 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2338 p
->policy
= SCHED_NORMAL
;
2339 p
->static_prio
= NICE_TO_PRIO(0);
2341 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2342 p
->static_prio
= NICE_TO_PRIO(0);
2344 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2345 set_load_weight(p
, false);
2348 * We don't need the reset flag anymore after the fork. It has
2349 * fulfilled its duty:
2351 p
->sched_reset_on_fork
= 0;
2354 if (dl_prio(p
->prio
)) {
2357 } else if (rt_prio(p
->prio
)) {
2358 p
->sched_class
= &rt_sched_class
;
2360 p
->sched_class
= &fair_sched_class
;
2363 init_entity_runnable_average(&p
->se
);
2366 * The child is not yet in the pid-hash so no cgroup attach races,
2367 * and the cgroup is pinned to this child due to cgroup_fork()
2368 * is ran before sched_fork().
2370 * Silence PROVE_RCU.
2372 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2374 * We're setting the CPU for the first time, we don't migrate,
2375 * so use __set_task_cpu().
2377 __set_task_cpu(p
, cpu
);
2378 if (p
->sched_class
->task_fork
)
2379 p
->sched_class
->task_fork(p
);
2380 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2382 #ifdef CONFIG_SCHED_INFO
2383 if (likely(sched_info_on()))
2384 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2386 #if defined(CONFIG_SMP)
2389 init_task_preempt_count(p
);
2391 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2392 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2399 unsigned long to_ratio(u64 period
, u64 runtime
)
2401 if (runtime
== RUNTIME_INF
)
2405 * Doing this here saves a lot of checks in all
2406 * the calling paths, and returning zero seems
2407 * safe for them anyway.
2412 return div64_u64(runtime
<< BW_SHIFT
, period
);
2416 * wake_up_new_task - wake up a newly created task for the first time.
2418 * This function will do some initial scheduler statistics housekeeping
2419 * that must be done for every newly created context, then puts the task
2420 * on the runqueue and wakes it.
2422 void wake_up_new_task(struct task_struct
*p
)
2427 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2428 p
->state
= TASK_RUNNING
;
2431 * Fork balancing, do it here and not earlier because:
2432 * - cpus_allowed can change in the fork path
2433 * - any previously selected CPU might disappear through hotplug
2435 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2436 * as we're not fully set-up yet.
2438 p
->recent_used_cpu
= task_cpu(p
);
2439 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2441 rq
= __task_rq_lock(p
, &rf
);
2442 update_rq_clock(rq
);
2443 post_init_entity_util_avg(&p
->se
);
2445 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
2446 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2447 trace_sched_wakeup_new(p
);
2448 check_preempt_curr(rq
, p
, WF_FORK
);
2450 if (p
->sched_class
->task_woken
) {
2452 * Nothing relies on rq->lock after this, so its fine to
2455 rq_unpin_lock(rq
, &rf
);
2456 p
->sched_class
->task_woken(rq
, p
);
2457 rq_repin_lock(rq
, &rf
);
2460 task_rq_unlock(rq
, p
, &rf
);
2463 #ifdef CONFIG_PREEMPT_NOTIFIERS
2465 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
2467 void preempt_notifier_inc(void)
2469 static_branch_inc(&preempt_notifier_key
);
2471 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2473 void preempt_notifier_dec(void)
2475 static_branch_dec(&preempt_notifier_key
);
2477 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2480 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2481 * @notifier: notifier struct to register
2483 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2485 if (!static_branch_unlikely(&preempt_notifier_key
))
2486 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2488 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2490 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2493 * preempt_notifier_unregister - no longer interested in preemption notifications
2494 * @notifier: notifier struct to unregister
2496 * This is *not* safe to call from within a preemption notifier.
2498 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2500 hlist_del(¬ifier
->link
);
2502 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2504 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2506 struct preempt_notifier
*notifier
;
2508 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2509 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2512 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2514 if (static_branch_unlikely(&preempt_notifier_key
))
2515 __fire_sched_in_preempt_notifiers(curr
);
2519 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2520 struct task_struct
*next
)
2522 struct preempt_notifier
*notifier
;
2524 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2525 notifier
->ops
->sched_out(notifier
, next
);
2528 static __always_inline
void
2529 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2530 struct task_struct
*next
)
2532 if (static_branch_unlikely(&preempt_notifier_key
))
2533 __fire_sched_out_preempt_notifiers(curr
, next
);
2536 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2538 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2543 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2544 struct task_struct
*next
)
2548 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2550 static inline void prepare_task(struct task_struct
*next
)
2554 * Claim the task as running, we do this before switching to it
2555 * such that any running task will have this set.
2561 static inline void finish_task(struct task_struct
*prev
)
2565 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2566 * We must ensure this doesn't happen until the switch is completely
2569 * In particular, the load of prev->state in finish_task_switch() must
2570 * happen before this.
2572 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2574 smp_store_release(&prev
->on_cpu
, 0);
2579 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
2582 * Since the runqueue lock will be released by the next
2583 * task (which is an invalid locking op but in the case
2584 * of the scheduler it's an obvious special-case), so we
2585 * do an early lockdep release here:
2587 rq_unpin_lock(rq
, rf
);
2588 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2589 #ifdef CONFIG_DEBUG_SPINLOCK
2590 /* this is a valid case when another task releases the spinlock */
2591 rq
->lock
.owner
= next
;
2595 static inline void finish_lock_switch(struct rq
*rq
)
2598 * If we are tracking spinlock dependencies then we have to
2599 * fix up the runqueue lock - which gets 'carried over' from
2600 * prev into current:
2602 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
2603 raw_spin_unlock_irq(&rq
->lock
);
2607 * NOP if the arch has not defined these:
2610 #ifndef prepare_arch_switch
2611 # define prepare_arch_switch(next) do { } while (0)
2614 #ifndef finish_arch_post_lock_switch
2615 # define finish_arch_post_lock_switch() do { } while (0)
2619 * prepare_task_switch - prepare to switch tasks
2620 * @rq: the runqueue preparing to switch
2621 * @prev: the current task that is being switched out
2622 * @next: the task we are going to switch to.
2624 * This is called with the rq lock held and interrupts off. It must
2625 * be paired with a subsequent finish_task_switch after the context
2628 * prepare_task_switch sets up locking and calls architecture specific
2632 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2633 struct task_struct
*next
)
2635 sched_info_switch(rq
, prev
, next
);
2636 perf_event_task_sched_out(prev
, next
);
2637 fire_sched_out_preempt_notifiers(prev
, next
);
2639 prepare_arch_switch(next
);
2643 * finish_task_switch - clean up after a task-switch
2644 * @prev: the thread we just switched away from.
2646 * finish_task_switch must be called after the context switch, paired
2647 * with a prepare_task_switch call before the context switch.
2648 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2649 * and do any other architecture-specific cleanup actions.
2651 * Note that we may have delayed dropping an mm in context_switch(). If
2652 * so, we finish that here outside of the runqueue lock. (Doing it
2653 * with the lock held can cause deadlocks; see schedule() for
2656 * The context switch have flipped the stack from under us and restored the
2657 * local variables which were saved when this task called schedule() in the
2658 * past. prev == current is still correct but we need to recalculate this_rq
2659 * because prev may have moved to another CPU.
2661 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2662 __releases(rq
->lock
)
2664 struct rq
*rq
= this_rq();
2665 struct mm_struct
*mm
= rq
->prev_mm
;
2669 * The previous task will have left us with a preempt_count of 2
2670 * because it left us after:
2673 * preempt_disable(); // 1
2675 * raw_spin_lock_irq(&rq->lock) // 2
2677 * Also, see FORK_PREEMPT_COUNT.
2679 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2680 "corrupted preempt_count: %s/%d/0x%x\n",
2681 current
->comm
, current
->pid
, preempt_count()))
2682 preempt_count_set(FORK_PREEMPT_COUNT
);
2687 * A task struct has one reference for the use as "current".
2688 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2689 * schedule one last time. The schedule call will never return, and
2690 * the scheduled task must drop that reference.
2692 * We must observe prev->state before clearing prev->on_cpu (in
2693 * finish_task), otherwise a concurrent wakeup can get prev
2694 * running on another CPU and we could rave with its RUNNING -> DEAD
2695 * transition, resulting in a double drop.
2697 prev_state
= prev
->state
;
2698 vtime_task_switch(prev
);
2699 perf_event_task_sched_in(prev
, current
);
2701 finish_lock_switch(rq
);
2702 finish_arch_post_lock_switch();
2704 fire_sched_in_preempt_notifiers(current
);
2706 * When switching through a kernel thread, the loop in
2707 * membarrier_{private,global}_expedited() may have observed that
2708 * kernel thread and not issued an IPI. It is therefore possible to
2709 * schedule between user->kernel->user threads without passing though
2710 * switch_mm(). Membarrier requires a barrier after storing to
2711 * rq->curr, before returning to userspace, so provide them here:
2713 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2714 * provided by mmdrop(),
2715 * - a sync_core for SYNC_CORE.
2718 membarrier_mm_sync_core_before_usermode(mm
);
2721 if (unlikely(prev_state
== TASK_DEAD
)) {
2722 if (prev
->sched_class
->task_dead
)
2723 prev
->sched_class
->task_dead(prev
);
2726 * Remove function-return probe instances associated with this
2727 * task and put them back on the free list.
2729 kprobe_flush_task(prev
);
2731 /* Task is done with its stack. */
2732 put_task_stack(prev
);
2734 put_task_struct(prev
);
2737 tick_nohz_task_switch();
2743 /* rq->lock is NOT held, but preemption is disabled */
2744 static void __balance_callback(struct rq
*rq
)
2746 struct callback_head
*head
, *next
;
2747 void (*func
)(struct rq
*rq
);
2748 unsigned long flags
;
2750 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2751 head
= rq
->balance_callback
;
2752 rq
->balance_callback
= NULL
;
2754 func
= (void (*)(struct rq
*))head
->func
;
2761 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2764 static inline void balance_callback(struct rq
*rq
)
2766 if (unlikely(rq
->balance_callback
))
2767 __balance_callback(rq
);
2772 static inline void balance_callback(struct rq
*rq
)
2779 * schedule_tail - first thing a freshly forked thread must call.
2780 * @prev: the thread we just switched away from.
2782 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2783 __releases(rq
->lock
)
2788 * New tasks start with FORK_PREEMPT_COUNT, see there and
2789 * finish_task_switch() for details.
2791 * finish_task_switch() will drop rq->lock() and lower preempt_count
2792 * and the preempt_enable() will end up enabling preemption (on
2793 * PREEMPT_COUNT kernels).
2796 rq
= finish_task_switch(prev
);
2797 balance_callback(rq
);
2800 if (current
->set_child_tid
)
2801 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2805 * context_switch - switch to the new MM and the new thread's register state.
2807 static __always_inline
struct rq
*
2808 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2809 struct task_struct
*next
, struct rq_flags
*rf
)
2811 struct mm_struct
*mm
, *oldmm
;
2813 prepare_task_switch(rq
, prev
, next
);
2816 oldmm
= prev
->active_mm
;
2818 * For paravirt, this is coupled with an exit in switch_to to
2819 * combine the page table reload and the switch backend into
2822 arch_start_context_switch(prev
);
2825 * If mm is non-NULL, we pass through switch_mm(). If mm is
2826 * NULL, we will pass through mmdrop() in finish_task_switch().
2827 * Both of these contain the full memory barrier required by
2828 * membarrier after storing to rq->curr, before returning to
2832 next
->active_mm
= oldmm
;
2834 enter_lazy_tlb(oldmm
, next
);
2836 switch_mm_irqs_off(oldmm
, mm
, next
);
2839 prev
->active_mm
= NULL
;
2840 rq
->prev_mm
= oldmm
;
2843 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
2845 prepare_lock_switch(rq
, next
, rf
);
2847 /* Here we just switch the register state and the stack. */
2848 switch_to(prev
, next
, prev
);
2851 return finish_task_switch(prev
);
2855 * nr_running and nr_context_switches:
2857 * externally visible scheduler statistics: current number of runnable
2858 * threads, total number of context switches performed since bootup.
2860 unsigned long nr_running(void)
2862 unsigned long i
, sum
= 0;
2864 for_each_online_cpu(i
)
2865 sum
+= cpu_rq(i
)->nr_running
;
2871 * Check if only the current task is running on the CPU.
2873 * Caution: this function does not check that the caller has disabled
2874 * preemption, thus the result might have a time-of-check-to-time-of-use
2875 * race. The caller is responsible to use it correctly, for example:
2877 * - from a non-preemptable section (of course)
2879 * - from a thread that is bound to a single CPU
2881 * - in a loop with very short iterations (e.g. a polling loop)
2883 bool single_task_running(void)
2885 return raw_rq()->nr_running
== 1;
2887 EXPORT_SYMBOL(single_task_running
);
2889 unsigned long long nr_context_switches(void)
2892 unsigned long long sum
= 0;
2894 for_each_possible_cpu(i
)
2895 sum
+= cpu_rq(i
)->nr_switches
;
2901 * IO-wait accounting, and how its mostly bollocks (on SMP).
2903 * The idea behind IO-wait account is to account the idle time that we could
2904 * have spend running if it were not for IO. That is, if we were to improve the
2905 * storage performance, we'd have a proportional reduction in IO-wait time.
2907 * This all works nicely on UP, where, when a task blocks on IO, we account
2908 * idle time as IO-wait, because if the storage were faster, it could've been
2909 * running and we'd not be idle.
2911 * This has been extended to SMP, by doing the same for each CPU. This however
2914 * Imagine for instance the case where two tasks block on one CPU, only the one
2915 * CPU will have IO-wait accounted, while the other has regular idle. Even
2916 * though, if the storage were faster, both could've ran at the same time,
2917 * utilising both CPUs.
2919 * This means, that when looking globally, the current IO-wait accounting on
2920 * SMP is a lower bound, by reason of under accounting.
2922 * Worse, since the numbers are provided per CPU, they are sometimes
2923 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2924 * associated with any one particular CPU, it can wake to another CPU than it
2925 * blocked on. This means the per CPU IO-wait number is meaningless.
2927 * Task CPU affinities can make all that even more 'interesting'.
2930 unsigned long nr_iowait(void)
2932 unsigned long i
, sum
= 0;
2934 for_each_possible_cpu(i
)
2935 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2941 * Consumers of these two interfaces, like for example the cpufreq menu
2942 * governor are using nonsensical data. Boosting frequency for a CPU that has
2943 * IO-wait which might not even end up running the task when it does become
2947 unsigned long nr_iowait_cpu(int cpu
)
2949 struct rq
*this = cpu_rq(cpu
);
2950 return atomic_read(&this->nr_iowait
);
2953 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2955 struct rq
*rq
= this_rq();
2956 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2957 *load
= rq
->load
.weight
;
2963 * sched_exec - execve() is a valuable balancing opportunity, because at
2964 * this point the task has the smallest effective memory and cache footprint.
2966 void sched_exec(void)
2968 struct task_struct
*p
= current
;
2969 unsigned long flags
;
2972 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2973 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2974 if (dest_cpu
== smp_processor_id())
2977 if (likely(cpu_active(dest_cpu
))) {
2978 struct migration_arg arg
= { p
, dest_cpu
};
2980 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2981 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2985 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2990 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2991 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2993 EXPORT_PER_CPU_SYMBOL(kstat
);
2994 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2997 * The function fair_sched_class.update_curr accesses the struct curr
2998 * and its field curr->exec_start; when called from task_sched_runtime(),
2999 * we observe a high rate of cache misses in practice.
3000 * Prefetching this data results in improved performance.
3002 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3004 #ifdef CONFIG_FAIR_GROUP_SCHED
3005 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3007 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3010 prefetch(&curr
->exec_start
);
3014 * Return accounted runtime for the task.
3015 * In case the task is currently running, return the runtime plus current's
3016 * pending runtime that have not been accounted yet.
3018 unsigned long long task_sched_runtime(struct task_struct
*p
)
3024 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3026 * 64-bit doesn't need locks to atomically read a 64-bit value.
3027 * So we have a optimization chance when the task's delta_exec is 0.
3028 * Reading ->on_cpu is racy, but this is ok.
3030 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3031 * If we race with it entering CPU, unaccounted time is 0. This is
3032 * indistinguishable from the read occurring a few cycles earlier.
3033 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3034 * been accounted, so we're correct here as well.
3036 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3037 return p
->se
.sum_exec_runtime
;
3040 rq
= task_rq_lock(p
, &rf
);
3042 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3043 * project cycles that may never be accounted to this
3044 * thread, breaking clock_gettime().
3046 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3047 prefetch_curr_exec_start(p
);
3048 update_rq_clock(rq
);
3049 p
->sched_class
->update_curr(rq
);
3051 ns
= p
->se
.sum_exec_runtime
;
3052 task_rq_unlock(rq
, p
, &rf
);
3058 * This function gets called by the timer code, with HZ frequency.
3059 * We call it with interrupts disabled.
3061 void scheduler_tick(void)
3063 int cpu
= smp_processor_id();
3064 struct rq
*rq
= cpu_rq(cpu
);
3065 struct task_struct
*curr
= rq
->curr
;
3072 update_rq_clock(rq
);
3073 curr
->sched_class
->task_tick(rq
, curr
, 0);
3074 cpu_load_update_active(rq
);
3075 calc_global_load_tick(rq
);
3079 perf_event_task_tick();
3082 rq
->idle_balance
= idle_cpu(cpu
);
3083 trigger_load_balance(rq
);
3087 #ifdef CONFIG_NO_HZ_FULL
3091 struct delayed_work work
;
3094 static struct tick_work __percpu
*tick_work_cpu
;
3096 static void sched_tick_remote(struct work_struct
*work
)
3098 struct delayed_work
*dwork
= to_delayed_work(work
);
3099 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
3100 int cpu
= twork
->cpu
;
3101 struct rq
*rq
= cpu_rq(cpu
);
3105 * Handle the tick only if it appears the remote CPU is running in full
3106 * dynticks mode. The check is racy by nature, but missing a tick or
3107 * having one too much is no big deal because the scheduler tick updates
3108 * statistics and checks timeslices in a time-independent way, regardless
3109 * of when exactly it is running.
3111 if (!idle_cpu(cpu
) && tick_nohz_tick_stopped_cpu(cpu
)) {
3112 struct task_struct
*curr
;
3115 rq_lock_irq(rq
, &rf
);
3116 update_rq_clock(rq
);
3118 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
3121 * Make sure the next tick runs within a reasonable
3124 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
3125 curr
->sched_class
->task_tick(rq
, curr
, 0);
3126 rq_unlock_irq(rq
, &rf
);
3130 * Run the remote tick once per second (1Hz). This arbitrary
3131 * frequency is large enough to avoid overload but short enough
3132 * to keep scheduler internal stats reasonably up to date.
3134 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
3137 static void sched_tick_start(int cpu
)
3139 struct tick_work
*twork
;
3141 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3144 WARN_ON_ONCE(!tick_work_cpu
);
3146 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3148 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
3149 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
3152 #ifdef CONFIG_HOTPLUG_CPU
3153 static void sched_tick_stop(int cpu
)
3155 struct tick_work
*twork
;
3157 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3160 WARN_ON_ONCE(!tick_work_cpu
);
3162 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3163 cancel_delayed_work_sync(&twork
->work
);
3165 #endif /* CONFIG_HOTPLUG_CPU */
3167 int __init
sched_tick_offload_init(void)
3169 tick_work_cpu
= alloc_percpu(struct tick_work
);
3170 BUG_ON(!tick_work_cpu
);
3175 #else /* !CONFIG_NO_HZ_FULL */
3176 static inline void sched_tick_start(int cpu
) { }
3177 static inline void sched_tick_stop(int cpu
) { }
3180 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3181 defined(CONFIG_PREEMPT_TRACER))
3183 * If the value passed in is equal to the current preempt count
3184 * then we just disabled preemption. Start timing the latency.
3186 static inline void preempt_latency_start(int val
)
3188 if (preempt_count() == val
) {
3189 unsigned long ip
= get_lock_parent_ip();
3190 #ifdef CONFIG_DEBUG_PREEMPT
3191 current
->preempt_disable_ip
= ip
;
3193 trace_preempt_off(CALLER_ADDR0
, ip
);
3197 void preempt_count_add(int val
)
3199 #ifdef CONFIG_DEBUG_PREEMPT
3203 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3206 __preempt_count_add(val
);
3207 #ifdef CONFIG_DEBUG_PREEMPT
3209 * Spinlock count overflowing soon?
3211 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3214 preempt_latency_start(val
);
3216 EXPORT_SYMBOL(preempt_count_add
);
3217 NOKPROBE_SYMBOL(preempt_count_add
);
3220 * If the value passed in equals to the current preempt count
3221 * then we just enabled preemption. Stop timing the latency.
3223 static inline void preempt_latency_stop(int val
)
3225 if (preempt_count() == val
)
3226 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3229 void preempt_count_sub(int val
)
3231 #ifdef CONFIG_DEBUG_PREEMPT
3235 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3238 * Is the spinlock portion underflowing?
3240 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3241 !(preempt_count() & PREEMPT_MASK
)))
3245 preempt_latency_stop(val
);
3246 __preempt_count_sub(val
);
3248 EXPORT_SYMBOL(preempt_count_sub
);
3249 NOKPROBE_SYMBOL(preempt_count_sub
);
3252 static inline void preempt_latency_start(int val
) { }
3253 static inline void preempt_latency_stop(int val
) { }
3256 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3258 #ifdef CONFIG_DEBUG_PREEMPT
3259 return p
->preempt_disable_ip
;
3266 * Print scheduling while atomic bug:
3268 static noinline
void __schedule_bug(struct task_struct
*prev
)
3270 /* Save this before calling printk(), since that will clobber it */
3271 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3273 if (oops_in_progress
)
3276 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3277 prev
->comm
, prev
->pid
, preempt_count());
3279 debug_show_held_locks(prev
);
3281 if (irqs_disabled())
3282 print_irqtrace_events(prev
);
3283 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3284 && in_atomic_preempt_off()) {
3285 pr_err("Preemption disabled at:");
3286 print_ip_sym(preempt_disable_ip
);
3290 panic("scheduling while atomic\n");
3293 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3297 * Various schedule()-time debugging checks and statistics:
3299 static inline void schedule_debug(struct task_struct
*prev
)
3301 #ifdef CONFIG_SCHED_STACK_END_CHECK
3302 if (task_stack_end_corrupted(prev
))
3303 panic("corrupted stack end detected inside scheduler\n");
3306 if (unlikely(in_atomic_preempt_off())) {
3307 __schedule_bug(prev
);
3308 preempt_count_set(PREEMPT_DISABLED
);
3312 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3314 schedstat_inc(this_rq()->sched_count
);
3318 * Pick up the highest-prio task:
3320 static inline struct task_struct
*
3321 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3323 const struct sched_class
*class;
3324 struct task_struct
*p
;
3327 * Optimization: we know that if all tasks are in the fair class we can
3328 * call that function directly, but only if the @prev task wasn't of a
3329 * higher scheduling class, because otherwise those loose the
3330 * opportunity to pull in more work from other CPUs.
3332 if (likely((prev
->sched_class
== &idle_sched_class
||
3333 prev
->sched_class
== &fair_sched_class
) &&
3334 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3336 p
= fair_sched_class
.pick_next_task(rq
, prev
, rf
);
3337 if (unlikely(p
== RETRY_TASK
))
3340 /* Assumes fair_sched_class->next == idle_sched_class */
3342 p
= idle_sched_class
.pick_next_task(rq
, prev
, rf
);
3348 for_each_class(class) {
3349 p
= class->pick_next_task(rq
, prev
, rf
);
3351 if (unlikely(p
== RETRY_TASK
))
3357 /* The idle class should always have a runnable task: */
3362 * __schedule() is the main scheduler function.
3364 * The main means of driving the scheduler and thus entering this function are:
3366 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3368 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3369 * paths. For example, see arch/x86/entry_64.S.
3371 * To drive preemption between tasks, the scheduler sets the flag in timer
3372 * interrupt handler scheduler_tick().
3374 * 3. Wakeups don't really cause entry into schedule(). They add a
3375 * task to the run-queue and that's it.
3377 * Now, if the new task added to the run-queue preempts the current
3378 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3379 * called on the nearest possible occasion:
3381 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3383 * - in syscall or exception context, at the next outmost
3384 * preempt_enable(). (this might be as soon as the wake_up()'s
3387 * - in IRQ context, return from interrupt-handler to
3388 * preemptible context
3390 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3393 * - cond_resched() call
3394 * - explicit schedule() call
3395 * - return from syscall or exception to user-space
3396 * - return from interrupt-handler to user-space
3398 * WARNING: must be called with preemption disabled!
3400 static void __sched notrace
__schedule(bool preempt
)
3402 struct task_struct
*prev
, *next
;
3403 unsigned long *switch_count
;
3408 cpu
= smp_processor_id();
3412 schedule_debug(prev
);
3414 if (sched_feat(HRTICK
))
3417 local_irq_disable();
3418 rcu_note_context_switch(preempt
);
3421 * Make sure that signal_pending_state()->signal_pending() below
3422 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3423 * done by the caller to avoid the race with signal_wake_up().
3425 * The membarrier system call requires a full memory barrier
3426 * after coming from user-space, before storing to rq->curr.
3429 smp_mb__after_spinlock();
3431 /* Promote REQ to ACT */
3432 rq
->clock_update_flags
<<= 1;
3433 update_rq_clock(rq
);
3435 switch_count
= &prev
->nivcsw
;
3436 if (!preempt
&& prev
->state
) {
3437 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3438 prev
->state
= TASK_RUNNING
;
3440 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
3443 if (prev
->in_iowait
) {
3444 atomic_inc(&rq
->nr_iowait
);
3445 delayacct_blkio_start();
3449 * If a worker went to sleep, notify and ask workqueue
3450 * whether it wants to wake up a task to maintain
3453 if (prev
->flags
& PF_WQ_WORKER
) {
3454 struct task_struct
*to_wakeup
;
3456 to_wakeup
= wq_worker_sleeping(prev
);
3458 try_to_wake_up_local(to_wakeup
, &rf
);
3461 switch_count
= &prev
->nvcsw
;
3464 next
= pick_next_task(rq
, prev
, &rf
);
3465 clear_tsk_need_resched(prev
);
3466 clear_preempt_need_resched();
3468 if (likely(prev
!= next
)) {
3472 * The membarrier system call requires each architecture
3473 * to have a full memory barrier after updating
3474 * rq->curr, before returning to user-space.
3476 * Here are the schemes providing that barrier on the
3477 * various architectures:
3478 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3479 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3480 * - finish_lock_switch() for weakly-ordered
3481 * architectures where spin_unlock is a full barrier,
3482 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3483 * is a RELEASE barrier),
3487 trace_sched_switch(preempt
, prev
, next
);
3489 /* Also unlocks the rq: */
3490 rq
= context_switch(rq
, prev
, next
, &rf
);
3492 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3493 rq_unlock_irq(rq
, &rf
);
3496 balance_callback(rq
);
3499 void __noreturn
do_task_dead(void)
3502 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3503 * when the following two conditions become true.
3504 * - There is race condition of mmap_sem (It is acquired by
3506 * - SMI occurs before setting TASK_RUNINNG.
3507 * (or hypervisor of virtual machine switches to other guest)
3508 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3510 * To avoid it, we have to wait for releasing tsk->pi_lock which
3511 * is held by try_to_wake_up()
3513 raw_spin_lock_irq(¤t
->pi_lock
);
3514 raw_spin_unlock_irq(¤t
->pi_lock
);
3516 /* Causes final put_task_struct in finish_task_switch(): */
3517 __set_current_state(TASK_DEAD
);
3519 /* Tell freezer to ignore us: */
3520 current
->flags
|= PF_NOFREEZE
;
3525 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3530 static inline void sched_submit_work(struct task_struct
*tsk
)
3532 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3535 * If we are going to sleep and we have plugged IO queued,
3536 * make sure to submit it to avoid deadlocks.
3538 if (blk_needs_flush_plug(tsk
))
3539 blk_schedule_flush_plug(tsk
);
3542 asmlinkage __visible
void __sched
schedule(void)
3544 struct task_struct
*tsk
= current
;
3546 sched_submit_work(tsk
);
3550 sched_preempt_enable_no_resched();
3551 } while (need_resched());
3553 EXPORT_SYMBOL(schedule
);
3556 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3557 * state (have scheduled out non-voluntarily) by making sure that all
3558 * tasks have either left the run queue or have gone into user space.
3559 * As idle tasks do not do either, they must not ever be preempted
3560 * (schedule out non-voluntarily).
3562 * schedule_idle() is similar to schedule_preempt_disable() except that it
3563 * never enables preemption because it does not call sched_submit_work().
3565 void __sched
schedule_idle(void)
3568 * As this skips calling sched_submit_work(), which the idle task does
3569 * regardless because that function is a nop when the task is in a
3570 * TASK_RUNNING state, make sure this isn't used someplace that the
3571 * current task can be in any other state. Note, idle is always in the
3572 * TASK_RUNNING state.
3574 WARN_ON_ONCE(current
->state
);
3577 } while (need_resched());
3580 #ifdef CONFIG_CONTEXT_TRACKING
3581 asmlinkage __visible
void __sched
schedule_user(void)
3584 * If we come here after a random call to set_need_resched(),
3585 * or we have been woken up remotely but the IPI has not yet arrived,
3586 * we haven't yet exited the RCU idle mode. Do it here manually until
3587 * we find a better solution.
3589 * NB: There are buggy callers of this function. Ideally we
3590 * should warn if prev_state != CONTEXT_USER, but that will trigger
3591 * too frequently to make sense yet.
3593 enum ctx_state prev_state
= exception_enter();
3595 exception_exit(prev_state
);
3600 * schedule_preempt_disabled - called with preemption disabled
3602 * Returns with preemption disabled. Note: preempt_count must be 1
3604 void __sched
schedule_preempt_disabled(void)
3606 sched_preempt_enable_no_resched();
3611 static void __sched notrace
preempt_schedule_common(void)
3615 * Because the function tracer can trace preempt_count_sub()
3616 * and it also uses preempt_enable/disable_notrace(), if
3617 * NEED_RESCHED is set, the preempt_enable_notrace() called
3618 * by the function tracer will call this function again and
3619 * cause infinite recursion.
3621 * Preemption must be disabled here before the function
3622 * tracer can trace. Break up preempt_disable() into two
3623 * calls. One to disable preemption without fear of being
3624 * traced. The other to still record the preemption latency,
3625 * which can also be traced by the function tracer.
3627 preempt_disable_notrace();
3628 preempt_latency_start(1);
3630 preempt_latency_stop(1);
3631 preempt_enable_no_resched_notrace();
3634 * Check again in case we missed a preemption opportunity
3635 * between schedule and now.
3637 } while (need_resched());
3640 #ifdef CONFIG_PREEMPT
3642 * this is the entry point to schedule() from in-kernel preemption
3643 * off of preempt_enable. Kernel preemptions off return from interrupt
3644 * occur there and call schedule directly.
3646 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3649 * If there is a non-zero preempt_count or interrupts are disabled,
3650 * we do not want to preempt the current task. Just return..
3652 if (likely(!preemptible()))
3655 preempt_schedule_common();
3657 NOKPROBE_SYMBOL(preempt_schedule
);
3658 EXPORT_SYMBOL(preempt_schedule
);
3661 * preempt_schedule_notrace - preempt_schedule called by tracing
3663 * The tracing infrastructure uses preempt_enable_notrace to prevent
3664 * recursion and tracing preempt enabling caused by the tracing
3665 * infrastructure itself. But as tracing can happen in areas coming
3666 * from userspace or just about to enter userspace, a preempt enable
3667 * can occur before user_exit() is called. This will cause the scheduler
3668 * to be called when the system is still in usermode.
3670 * To prevent this, the preempt_enable_notrace will use this function
3671 * instead of preempt_schedule() to exit user context if needed before
3672 * calling the scheduler.
3674 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3676 enum ctx_state prev_ctx
;
3678 if (likely(!preemptible()))
3683 * Because the function tracer can trace preempt_count_sub()
3684 * and it also uses preempt_enable/disable_notrace(), if
3685 * NEED_RESCHED is set, the preempt_enable_notrace() called
3686 * by the function tracer will call this function again and
3687 * cause infinite recursion.
3689 * Preemption must be disabled here before the function
3690 * tracer can trace. Break up preempt_disable() into two
3691 * calls. One to disable preemption without fear of being
3692 * traced. The other to still record the preemption latency,
3693 * which can also be traced by the function tracer.
3695 preempt_disable_notrace();
3696 preempt_latency_start(1);
3698 * Needs preempt disabled in case user_exit() is traced
3699 * and the tracer calls preempt_enable_notrace() causing
3700 * an infinite recursion.
3702 prev_ctx
= exception_enter();
3704 exception_exit(prev_ctx
);
3706 preempt_latency_stop(1);
3707 preempt_enable_no_resched_notrace();
3708 } while (need_resched());
3710 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3712 #endif /* CONFIG_PREEMPT */
3715 * this is the entry point to schedule() from kernel preemption
3716 * off of irq context.
3717 * Note, that this is called and return with irqs disabled. This will
3718 * protect us against recursive calling from irq.
3720 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3722 enum ctx_state prev_state
;
3724 /* Catch callers which need to be fixed */
3725 BUG_ON(preempt_count() || !irqs_disabled());
3727 prev_state
= exception_enter();
3733 local_irq_disable();
3734 sched_preempt_enable_no_resched();
3735 } while (need_resched());
3737 exception_exit(prev_state
);
3740 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
3743 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3745 EXPORT_SYMBOL(default_wake_function
);
3747 #ifdef CONFIG_RT_MUTEXES
3749 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
3752 prio
= min(prio
, pi_task
->prio
);
3757 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3759 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3761 return __rt_effective_prio(pi_task
, prio
);
3765 * rt_mutex_setprio - set the current priority of a task
3767 * @pi_task: donor task
3769 * This function changes the 'effective' priority of a task. It does
3770 * not touch ->normal_prio like __setscheduler().
3772 * Used by the rt_mutex code to implement priority inheritance
3773 * logic. Call site only calls if the priority of the task changed.
3775 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
3777 int prio
, oldprio
, queued
, running
, queue_flag
=
3778 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
3779 const struct sched_class
*prev_class
;
3783 /* XXX used to be waiter->prio, not waiter->task->prio */
3784 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
3787 * If nothing changed; bail early.
3789 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
3792 rq
= __task_rq_lock(p
, &rf
);
3793 update_rq_clock(rq
);
3795 * Set under pi_lock && rq->lock, such that the value can be used under
3798 * Note that there is loads of tricky to make this pointer cache work
3799 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3800 * ensure a task is de-boosted (pi_task is set to NULL) before the
3801 * task is allowed to run again (and can exit). This ensures the pointer
3802 * points to a blocked task -- which guaratees the task is present.
3804 p
->pi_top_task
= pi_task
;
3807 * For FIFO/RR we only need to set prio, if that matches we're done.
3809 if (prio
== p
->prio
&& !dl_prio(prio
))
3813 * Idle task boosting is a nono in general. There is one
3814 * exception, when PREEMPT_RT and NOHZ is active:
3816 * The idle task calls get_next_timer_interrupt() and holds
3817 * the timer wheel base->lock on the CPU and another CPU wants
3818 * to access the timer (probably to cancel it). We can safely
3819 * ignore the boosting request, as the idle CPU runs this code
3820 * with interrupts disabled and will complete the lock
3821 * protected section without being interrupted. So there is no
3822 * real need to boost.
3824 if (unlikely(p
== rq
->idle
)) {
3825 WARN_ON(p
!= rq
->curr
);
3826 WARN_ON(p
->pi_blocked_on
);
3830 trace_sched_pi_setprio(p
, pi_task
);
3833 if (oldprio
== prio
)
3834 queue_flag
&= ~DEQUEUE_MOVE
;
3836 prev_class
= p
->sched_class
;
3837 queued
= task_on_rq_queued(p
);
3838 running
= task_current(rq
, p
);
3840 dequeue_task(rq
, p
, queue_flag
);
3842 put_prev_task(rq
, p
);
3845 * Boosting condition are:
3846 * 1. -rt task is running and holds mutex A
3847 * --> -dl task blocks on mutex A
3849 * 2. -dl task is running and holds mutex A
3850 * --> -dl task blocks on mutex A and could preempt the
3853 if (dl_prio(prio
)) {
3854 if (!dl_prio(p
->normal_prio
) ||
3855 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3856 p
->dl
.dl_boosted
= 1;
3857 queue_flag
|= ENQUEUE_REPLENISH
;
3859 p
->dl
.dl_boosted
= 0;
3860 p
->sched_class
= &dl_sched_class
;
3861 } else if (rt_prio(prio
)) {
3862 if (dl_prio(oldprio
))
3863 p
->dl
.dl_boosted
= 0;
3865 queue_flag
|= ENQUEUE_HEAD
;
3866 p
->sched_class
= &rt_sched_class
;
3868 if (dl_prio(oldprio
))
3869 p
->dl
.dl_boosted
= 0;
3870 if (rt_prio(oldprio
))
3872 p
->sched_class
= &fair_sched_class
;
3878 enqueue_task(rq
, p
, queue_flag
);
3880 set_curr_task(rq
, p
);
3882 check_class_changed(rq
, p
, prev_class
, oldprio
);
3884 /* Avoid rq from going away on us: */
3886 __task_rq_unlock(rq
, &rf
);
3888 balance_callback(rq
);
3892 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3898 void set_user_nice(struct task_struct
*p
, long nice
)
3900 bool queued
, running
;
3901 int old_prio
, delta
;
3905 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3908 * We have to be careful, if called from sys_setpriority(),
3909 * the task might be in the middle of scheduling on another CPU.
3911 rq
= task_rq_lock(p
, &rf
);
3912 update_rq_clock(rq
);
3915 * The RT priorities are set via sched_setscheduler(), but we still
3916 * allow the 'normal' nice value to be set - but as expected
3917 * it wont have any effect on scheduling until the task is
3918 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3920 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3921 p
->static_prio
= NICE_TO_PRIO(nice
);
3924 queued
= task_on_rq_queued(p
);
3925 running
= task_current(rq
, p
);
3927 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
3929 put_prev_task(rq
, p
);
3931 p
->static_prio
= NICE_TO_PRIO(nice
);
3932 set_load_weight(p
, true);
3934 p
->prio
= effective_prio(p
);
3935 delta
= p
->prio
- old_prio
;
3938 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
3940 * If the task increased its priority or is running and
3941 * lowered its priority, then reschedule its CPU:
3943 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3947 set_curr_task(rq
, p
);
3949 task_rq_unlock(rq
, p
, &rf
);
3951 EXPORT_SYMBOL(set_user_nice
);
3954 * can_nice - check if a task can reduce its nice value
3958 int can_nice(const struct task_struct
*p
, const int nice
)
3960 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3961 int nice_rlim
= nice_to_rlimit(nice
);
3963 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3964 capable(CAP_SYS_NICE
));
3967 #ifdef __ARCH_WANT_SYS_NICE
3970 * sys_nice - change the priority of the current process.
3971 * @increment: priority increment
3973 * sys_setpriority is a more generic, but much slower function that
3974 * does similar things.
3976 SYSCALL_DEFINE1(nice
, int, increment
)
3981 * Setpriority might change our priority at the same moment.
3982 * We don't have to worry. Conceptually one call occurs first
3983 * and we have a single winner.
3985 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3986 nice
= task_nice(current
) + increment
;
3988 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3989 if (increment
< 0 && !can_nice(current
, nice
))
3992 retval
= security_task_setnice(current
, nice
);
3996 set_user_nice(current
, nice
);
4003 * task_prio - return the priority value of a given task.
4004 * @p: the task in question.
4006 * Return: The priority value as seen by users in /proc.
4007 * RT tasks are offset by -200. Normal tasks are centered
4008 * around 0, value goes from -16 to +15.
4010 int task_prio(const struct task_struct
*p
)
4012 return p
->prio
- MAX_RT_PRIO
;
4016 * idle_cpu - is a given CPU idle currently?
4017 * @cpu: the processor in question.
4019 * Return: 1 if the CPU is currently idle. 0 otherwise.
4021 int idle_cpu(int cpu
)
4023 struct rq
*rq
= cpu_rq(cpu
);
4025 if (rq
->curr
!= rq
->idle
)
4032 if (!llist_empty(&rq
->wake_list
))
4040 * idle_task - return the idle task for a given CPU.
4041 * @cpu: the processor in question.
4043 * Return: The idle task for the CPU @cpu.
4045 struct task_struct
*idle_task(int cpu
)
4047 return cpu_rq(cpu
)->idle
;
4051 * find_process_by_pid - find a process with a matching PID value.
4052 * @pid: the pid in question.
4054 * The task of @pid, if found. %NULL otherwise.
4056 static struct task_struct
*find_process_by_pid(pid_t pid
)
4058 return pid
? find_task_by_vpid(pid
) : current
;
4062 * sched_setparam() passes in -1 for its policy, to let the functions
4063 * it calls know not to change it.
4065 #define SETPARAM_POLICY -1
4067 static void __setscheduler_params(struct task_struct
*p
,
4068 const struct sched_attr
*attr
)
4070 int policy
= attr
->sched_policy
;
4072 if (policy
== SETPARAM_POLICY
)
4077 if (dl_policy(policy
))
4078 __setparam_dl(p
, attr
);
4079 else if (fair_policy(policy
))
4080 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
4083 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4084 * !rt_policy. Always setting this ensures that things like
4085 * getparam()/getattr() don't report silly values for !rt tasks.
4087 p
->rt_priority
= attr
->sched_priority
;
4088 p
->normal_prio
= normal_prio(p
);
4089 set_load_weight(p
, true);
4092 /* Actually do priority change: must hold pi & rq lock. */
4093 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
4094 const struct sched_attr
*attr
, bool keep_boost
)
4096 __setscheduler_params(p
, attr
);
4099 * Keep a potential priority boosting if called from
4100 * sched_setscheduler().
4102 p
->prio
= normal_prio(p
);
4104 p
->prio
= rt_effective_prio(p
, p
->prio
);
4106 if (dl_prio(p
->prio
))
4107 p
->sched_class
= &dl_sched_class
;
4108 else if (rt_prio(p
->prio
))
4109 p
->sched_class
= &rt_sched_class
;
4111 p
->sched_class
= &fair_sched_class
;
4115 * Check the target process has a UID that matches the current process's:
4117 static bool check_same_owner(struct task_struct
*p
)
4119 const struct cred
*cred
= current_cred(), *pcred
;
4123 pcred
= __task_cred(p
);
4124 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4125 uid_eq(cred
->euid
, pcred
->uid
));
4130 static int __sched_setscheduler(struct task_struct
*p
,
4131 const struct sched_attr
*attr
,
4134 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4135 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4136 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4137 int new_effective_prio
, policy
= attr
->sched_policy
;
4138 const struct sched_class
*prev_class
;
4141 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4144 /* The pi code expects interrupts enabled */
4145 BUG_ON(pi
&& in_interrupt());
4147 /* Double check policy once rq lock held: */
4149 reset_on_fork
= p
->sched_reset_on_fork
;
4150 policy
= oldpolicy
= p
->policy
;
4152 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4154 if (!valid_policy(policy
))
4158 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
4162 * Valid priorities for SCHED_FIFO and SCHED_RR are
4163 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4164 * SCHED_BATCH and SCHED_IDLE is 0.
4166 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4167 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4169 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4170 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4174 * Allow unprivileged RT tasks to decrease priority:
4176 if (user
&& !capable(CAP_SYS_NICE
)) {
4177 if (fair_policy(policy
)) {
4178 if (attr
->sched_nice
< task_nice(p
) &&
4179 !can_nice(p
, attr
->sched_nice
))
4183 if (rt_policy(policy
)) {
4184 unsigned long rlim_rtprio
=
4185 task_rlimit(p
, RLIMIT_RTPRIO
);
4187 /* Can't set/change the rt policy: */
4188 if (policy
!= p
->policy
&& !rlim_rtprio
)
4191 /* Can't increase priority: */
4192 if (attr
->sched_priority
> p
->rt_priority
&&
4193 attr
->sched_priority
> rlim_rtprio
)
4198 * Can't set/change SCHED_DEADLINE policy at all for now
4199 * (safest behavior); in the future we would like to allow
4200 * unprivileged DL tasks to increase their relative deadline
4201 * or reduce their runtime (both ways reducing utilization)
4203 if (dl_policy(policy
))
4207 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4208 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4210 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4211 if (!can_nice(p
, task_nice(p
)))
4215 /* Can't change other user's priorities: */
4216 if (!check_same_owner(p
))
4219 /* Normal users shall not reset the sched_reset_on_fork flag: */
4220 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4225 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
4228 retval
= security_task_setscheduler(p
);
4234 * Make sure no PI-waiters arrive (or leave) while we are
4235 * changing the priority of the task:
4237 * To be able to change p->policy safely, the appropriate
4238 * runqueue lock must be held.
4240 rq
= task_rq_lock(p
, &rf
);
4241 update_rq_clock(rq
);
4244 * Changing the policy of the stop threads its a very bad idea:
4246 if (p
== rq
->stop
) {
4247 task_rq_unlock(rq
, p
, &rf
);
4252 * If not changing anything there's no need to proceed further,
4253 * but store a possible modification of reset_on_fork.
4255 if (unlikely(policy
== p
->policy
)) {
4256 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4258 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4260 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4263 p
->sched_reset_on_fork
= reset_on_fork
;
4264 task_rq_unlock(rq
, p
, &rf
);
4270 #ifdef CONFIG_RT_GROUP_SCHED
4272 * Do not allow realtime tasks into groups that have no runtime
4275 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4276 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4277 !task_group_is_autogroup(task_group(p
))) {
4278 task_rq_unlock(rq
, p
, &rf
);
4283 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
4284 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
4285 cpumask_t
*span
= rq
->rd
->span
;
4288 * Don't allow tasks with an affinity mask smaller than
4289 * the entire root_domain to become SCHED_DEADLINE. We
4290 * will also fail if there's no bandwidth available.
4292 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4293 rq
->rd
->dl_bw
.bw
== 0) {
4294 task_rq_unlock(rq
, p
, &rf
);
4301 /* Re-check policy now with rq lock held: */
4302 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4303 policy
= oldpolicy
= -1;
4304 task_rq_unlock(rq
, p
, &rf
);
4309 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4310 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4313 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
4314 task_rq_unlock(rq
, p
, &rf
);
4318 p
->sched_reset_on_fork
= reset_on_fork
;
4323 * Take priority boosted tasks into account. If the new
4324 * effective priority is unchanged, we just store the new
4325 * normal parameters and do not touch the scheduler class and
4326 * the runqueue. This will be done when the task deboost
4329 new_effective_prio
= rt_effective_prio(p
, newprio
);
4330 if (new_effective_prio
== oldprio
)
4331 queue_flags
&= ~DEQUEUE_MOVE
;
4334 queued
= task_on_rq_queued(p
);
4335 running
= task_current(rq
, p
);
4337 dequeue_task(rq
, p
, queue_flags
);
4339 put_prev_task(rq
, p
);
4341 prev_class
= p
->sched_class
;
4342 __setscheduler(rq
, p
, attr
, pi
);
4346 * We enqueue to tail when the priority of a task is
4347 * increased (user space view).
4349 if (oldprio
< p
->prio
)
4350 queue_flags
|= ENQUEUE_HEAD
;
4352 enqueue_task(rq
, p
, queue_flags
);
4355 set_curr_task(rq
, p
);
4357 check_class_changed(rq
, p
, prev_class
, oldprio
);
4359 /* Avoid rq from going away on us: */
4361 task_rq_unlock(rq
, p
, &rf
);
4364 rt_mutex_adjust_pi(p
);
4366 /* Run balance callbacks after we've adjusted the PI chain: */
4367 balance_callback(rq
);
4373 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4374 const struct sched_param
*param
, bool check
)
4376 struct sched_attr attr
= {
4377 .sched_policy
= policy
,
4378 .sched_priority
= param
->sched_priority
,
4379 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4382 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4383 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4384 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4385 policy
&= ~SCHED_RESET_ON_FORK
;
4386 attr
.sched_policy
= policy
;
4389 return __sched_setscheduler(p
, &attr
, check
, true);
4392 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4393 * @p: the task in question.
4394 * @policy: new policy.
4395 * @param: structure containing the new RT priority.
4397 * Return: 0 on success. An error code otherwise.
4399 * NOTE that the task may be already dead.
4401 int sched_setscheduler(struct task_struct
*p
, int policy
,
4402 const struct sched_param
*param
)
4404 return _sched_setscheduler(p
, policy
, param
, true);
4406 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4408 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4410 return __sched_setscheduler(p
, attr
, true, true);
4412 EXPORT_SYMBOL_GPL(sched_setattr
);
4414 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
4416 return __sched_setscheduler(p
, attr
, false, true);
4420 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4421 * @p: the task in question.
4422 * @policy: new policy.
4423 * @param: structure containing the new RT priority.
4425 * Just like sched_setscheduler, only don't bother checking if the
4426 * current context has permission. For example, this is needed in
4427 * stop_machine(): we create temporary high priority worker threads,
4428 * but our caller might not have that capability.
4430 * Return: 0 on success. An error code otherwise.
4432 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4433 const struct sched_param
*param
)
4435 return _sched_setscheduler(p
, policy
, param
, false);
4437 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4440 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4442 struct sched_param lparam
;
4443 struct task_struct
*p
;
4446 if (!param
|| pid
< 0)
4448 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4453 p
= find_process_by_pid(pid
);
4455 retval
= sched_setscheduler(p
, policy
, &lparam
);
4462 * Mimics kernel/events/core.c perf_copy_attr().
4464 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
4469 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4472 /* Zero the full structure, so that a short copy will be nice: */
4473 memset(attr
, 0, sizeof(*attr
));
4475 ret
= get_user(size
, &uattr
->size
);
4479 /* Bail out on silly large: */
4480 if (size
> PAGE_SIZE
)
4483 /* ABI compatibility quirk: */
4485 size
= SCHED_ATTR_SIZE_VER0
;
4487 if (size
< SCHED_ATTR_SIZE_VER0
)
4491 * If we're handed a bigger struct than we know of,
4492 * ensure all the unknown bits are 0 - i.e. new
4493 * user-space does not rely on any kernel feature
4494 * extensions we dont know about yet.
4496 if (size
> sizeof(*attr
)) {
4497 unsigned char __user
*addr
;
4498 unsigned char __user
*end
;
4501 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4502 end
= (void __user
*)uattr
+ size
;
4504 for (; addr
< end
; addr
++) {
4505 ret
= get_user(val
, addr
);
4511 size
= sizeof(*attr
);
4514 ret
= copy_from_user(attr
, uattr
, size
);
4519 * XXX: Do we want to be lenient like existing syscalls; or do we want
4520 * to be strict and return an error on out-of-bounds values?
4522 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4527 put_user(sizeof(*attr
), &uattr
->size
);
4532 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4533 * @pid: the pid in question.
4534 * @policy: new policy.
4535 * @param: structure containing the new RT priority.
4537 * Return: 0 on success. An error code otherwise.
4539 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
4544 return do_sched_setscheduler(pid
, policy
, param
);
4548 * sys_sched_setparam - set/change the RT priority of a thread
4549 * @pid: the pid in question.
4550 * @param: structure containing the new RT priority.
4552 * Return: 0 on success. An error code otherwise.
4554 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4556 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4560 * sys_sched_setattr - same as above, but with extended sched_attr
4561 * @pid: the pid in question.
4562 * @uattr: structure containing the extended parameters.
4563 * @flags: for future extension.
4565 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4566 unsigned int, flags
)
4568 struct sched_attr attr
;
4569 struct task_struct
*p
;
4572 if (!uattr
|| pid
< 0 || flags
)
4575 retval
= sched_copy_attr(uattr
, &attr
);
4579 if ((int)attr
.sched_policy
< 0)
4584 p
= find_process_by_pid(pid
);
4586 retval
= sched_setattr(p
, &attr
);
4593 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4594 * @pid: the pid in question.
4596 * Return: On success, the policy of the thread. Otherwise, a negative error
4599 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4601 struct task_struct
*p
;
4609 p
= find_process_by_pid(pid
);
4611 retval
= security_task_getscheduler(p
);
4614 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4621 * sys_sched_getparam - get the RT priority of a thread
4622 * @pid: the pid in question.
4623 * @param: structure containing the RT priority.
4625 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4628 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4630 struct sched_param lp
= { .sched_priority
= 0 };
4631 struct task_struct
*p
;
4634 if (!param
|| pid
< 0)
4638 p
= find_process_by_pid(pid
);
4643 retval
= security_task_getscheduler(p
);
4647 if (task_has_rt_policy(p
))
4648 lp
.sched_priority
= p
->rt_priority
;
4652 * This one might sleep, we cannot do it with a spinlock held ...
4654 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4663 static int sched_read_attr(struct sched_attr __user
*uattr
,
4664 struct sched_attr
*attr
,
4669 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4673 * If we're handed a smaller struct than we know of,
4674 * ensure all the unknown bits are 0 - i.e. old
4675 * user-space does not get uncomplete information.
4677 if (usize
< sizeof(*attr
)) {
4678 unsigned char *addr
;
4681 addr
= (void *)attr
+ usize
;
4682 end
= (void *)attr
+ sizeof(*attr
);
4684 for (; addr
< end
; addr
++) {
4692 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4700 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4701 * @pid: the pid in question.
4702 * @uattr: structure containing the extended parameters.
4703 * @size: sizeof(attr) for fwd/bwd comp.
4704 * @flags: for future extension.
4706 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4707 unsigned int, size
, unsigned int, flags
)
4709 struct sched_attr attr
= {
4710 .size
= sizeof(struct sched_attr
),
4712 struct task_struct
*p
;
4715 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4716 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4720 p
= find_process_by_pid(pid
);
4725 retval
= security_task_getscheduler(p
);
4729 attr
.sched_policy
= p
->policy
;
4730 if (p
->sched_reset_on_fork
)
4731 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4732 if (task_has_dl_policy(p
))
4733 __getparam_dl(p
, &attr
);
4734 else if (task_has_rt_policy(p
))
4735 attr
.sched_priority
= p
->rt_priority
;
4737 attr
.sched_nice
= task_nice(p
);
4741 retval
= sched_read_attr(uattr
, &attr
, size
);
4749 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4751 cpumask_var_t cpus_allowed
, new_mask
;
4752 struct task_struct
*p
;
4757 p
= find_process_by_pid(pid
);
4763 /* Prevent p going away */
4767 if (p
->flags
& PF_NO_SETAFFINITY
) {
4771 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4775 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4777 goto out_free_cpus_allowed
;
4780 if (!check_same_owner(p
)) {
4782 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4784 goto out_free_new_mask
;
4789 retval
= security_task_setscheduler(p
);
4791 goto out_free_new_mask
;
4794 cpuset_cpus_allowed(p
, cpus_allowed
);
4795 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4798 * Since bandwidth control happens on root_domain basis,
4799 * if admission test is enabled, we only admit -deadline
4800 * tasks allowed to run on all the CPUs in the task's
4804 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4806 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4809 goto out_free_new_mask
;
4815 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4818 cpuset_cpus_allowed(p
, cpus_allowed
);
4819 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4821 * We must have raced with a concurrent cpuset
4822 * update. Just reset the cpus_allowed to the
4823 * cpuset's cpus_allowed
4825 cpumask_copy(new_mask
, cpus_allowed
);
4830 free_cpumask_var(new_mask
);
4831 out_free_cpus_allowed
:
4832 free_cpumask_var(cpus_allowed
);
4838 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4839 struct cpumask
*new_mask
)
4841 if (len
< cpumask_size())
4842 cpumask_clear(new_mask
);
4843 else if (len
> cpumask_size())
4844 len
= cpumask_size();
4846 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4850 * sys_sched_setaffinity - set the CPU affinity of a process
4851 * @pid: pid of the process
4852 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4853 * @user_mask_ptr: user-space pointer to the new CPU mask
4855 * Return: 0 on success. An error code otherwise.
4857 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4858 unsigned long __user
*, user_mask_ptr
)
4860 cpumask_var_t new_mask
;
4863 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4866 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4868 retval
= sched_setaffinity(pid
, new_mask
);
4869 free_cpumask_var(new_mask
);
4873 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4875 struct task_struct
*p
;
4876 unsigned long flags
;
4882 p
= find_process_by_pid(pid
);
4886 retval
= security_task_getscheduler(p
);
4890 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4891 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4892 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4901 * sys_sched_getaffinity - get the CPU affinity of a process
4902 * @pid: pid of the process
4903 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4904 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4906 * Return: size of CPU mask copied to user_mask_ptr on success. An
4907 * error code otherwise.
4909 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4910 unsigned long __user
*, user_mask_ptr
)
4915 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4917 if (len
& (sizeof(unsigned long)-1))
4920 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4923 ret
= sched_getaffinity(pid
, mask
);
4925 unsigned int retlen
= min(len
, cpumask_size());
4927 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4932 free_cpumask_var(mask
);
4938 * sys_sched_yield - yield the current processor to other threads.
4940 * This function yields the current CPU to other tasks. If there are no
4941 * other threads running on this CPU then this function will return.
4945 static void do_sched_yield(void)
4950 local_irq_disable();
4954 schedstat_inc(rq
->yld_count
);
4955 current
->sched_class
->yield_task(rq
);
4958 * Since we are going to call schedule() anyway, there's
4959 * no need to preempt or enable interrupts:
4963 sched_preempt_enable_no_resched();
4968 SYSCALL_DEFINE0(sched_yield
)
4974 #ifndef CONFIG_PREEMPT
4975 int __sched
_cond_resched(void)
4977 if (should_resched(0)) {
4978 preempt_schedule_common();
4984 EXPORT_SYMBOL(_cond_resched
);
4988 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4989 * call schedule, and on return reacquire the lock.
4991 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4992 * operations here to prevent schedule() from being called twice (once via
4993 * spin_unlock(), once by hand).
4995 int __cond_resched_lock(spinlock_t
*lock
)
4997 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
5000 lockdep_assert_held(lock
);
5002 if (spin_needbreak(lock
) || resched
) {
5005 preempt_schedule_common();
5013 EXPORT_SYMBOL(__cond_resched_lock
);
5015 int __sched
__cond_resched_softirq(void)
5017 BUG_ON(!in_softirq());
5019 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
5021 preempt_schedule_common();
5027 EXPORT_SYMBOL(__cond_resched_softirq
);
5030 * yield - yield the current processor to other threads.
5032 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5034 * The scheduler is at all times free to pick the calling task as the most
5035 * eligible task to run, if removing the yield() call from your code breaks
5036 * it, its already broken.
5038 * Typical broken usage is:
5043 * where one assumes that yield() will let 'the other' process run that will
5044 * make event true. If the current task is a SCHED_FIFO task that will never
5045 * happen. Never use yield() as a progress guarantee!!
5047 * If you want to use yield() to wait for something, use wait_event().
5048 * If you want to use yield() to be 'nice' for others, use cond_resched().
5049 * If you still want to use yield(), do not!
5051 void __sched
yield(void)
5053 set_current_state(TASK_RUNNING
);
5056 EXPORT_SYMBOL(yield
);
5059 * yield_to - yield the current processor to another thread in
5060 * your thread group, or accelerate that thread toward the
5061 * processor it's on.
5063 * @preempt: whether task preemption is allowed or not
5065 * It's the caller's job to ensure that the target task struct
5066 * can't go away on us before we can do any checks.
5069 * true (>0) if we indeed boosted the target task.
5070 * false (0) if we failed to boost the target.
5071 * -ESRCH if there's no task to yield to.
5073 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5075 struct task_struct
*curr
= current
;
5076 struct rq
*rq
, *p_rq
;
5077 unsigned long flags
;
5080 local_irq_save(flags
);
5086 * If we're the only runnable task on the rq and target rq also
5087 * has only one task, there's absolutely no point in yielding.
5089 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5094 double_rq_lock(rq
, p_rq
);
5095 if (task_rq(p
) != p_rq
) {
5096 double_rq_unlock(rq
, p_rq
);
5100 if (!curr
->sched_class
->yield_to_task
)
5103 if (curr
->sched_class
!= p
->sched_class
)
5106 if (task_running(p_rq
, p
) || p
->state
)
5109 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5111 schedstat_inc(rq
->yld_count
);
5113 * Make p's CPU reschedule; pick_next_entity takes care of
5116 if (preempt
&& rq
!= p_rq
)
5121 double_rq_unlock(rq
, p_rq
);
5123 local_irq_restore(flags
);
5130 EXPORT_SYMBOL_GPL(yield_to
);
5132 int io_schedule_prepare(void)
5134 int old_iowait
= current
->in_iowait
;
5136 current
->in_iowait
= 1;
5137 blk_schedule_flush_plug(current
);
5142 void io_schedule_finish(int token
)
5144 current
->in_iowait
= token
;
5148 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5149 * that process accounting knows that this is a task in IO wait state.
5151 long __sched
io_schedule_timeout(long timeout
)
5156 token
= io_schedule_prepare();
5157 ret
= schedule_timeout(timeout
);
5158 io_schedule_finish(token
);
5162 EXPORT_SYMBOL(io_schedule_timeout
);
5164 void io_schedule(void)
5168 token
= io_schedule_prepare();
5170 io_schedule_finish(token
);
5172 EXPORT_SYMBOL(io_schedule
);
5175 * sys_sched_get_priority_max - return maximum RT priority.
5176 * @policy: scheduling class.
5178 * Return: On success, this syscall returns the maximum
5179 * rt_priority that can be used by a given scheduling class.
5180 * On failure, a negative error code is returned.
5182 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5189 ret
= MAX_USER_RT_PRIO
-1;
5191 case SCHED_DEADLINE
:
5202 * sys_sched_get_priority_min - return minimum RT priority.
5203 * @policy: scheduling class.
5205 * Return: On success, this syscall returns the minimum
5206 * rt_priority that can be used by a given scheduling class.
5207 * On failure, a negative error code is returned.
5209 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5218 case SCHED_DEADLINE
:
5227 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
5229 struct task_struct
*p
;
5230 unsigned int time_slice
;
5240 p
= find_process_by_pid(pid
);
5244 retval
= security_task_getscheduler(p
);
5248 rq
= task_rq_lock(p
, &rf
);
5250 if (p
->sched_class
->get_rr_interval
)
5251 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5252 task_rq_unlock(rq
, p
, &rf
);
5255 jiffies_to_timespec64(time_slice
, t
);
5264 * sys_sched_rr_get_interval - return the default timeslice of a process.
5265 * @pid: pid of the process.
5266 * @interval: userspace pointer to the timeslice value.
5268 * this syscall writes the default timeslice value of a given process
5269 * into the user-space timespec buffer. A value of '0' means infinity.
5271 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5274 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5275 struct timespec __user
*, interval
)
5277 struct timespec64 t
;
5278 int retval
= sched_rr_get_interval(pid
, &t
);
5281 retval
= put_timespec64(&t
, interval
);
5286 #ifdef CONFIG_COMPAT
5287 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval
,
5289 struct compat_timespec __user
*, interval
)
5291 struct timespec64 t
;
5292 int retval
= sched_rr_get_interval(pid
, &t
);
5295 retval
= compat_put_timespec64(&t
, interval
);
5300 void sched_show_task(struct task_struct
*p
)
5302 unsigned long free
= 0;
5305 if (!try_get_task_stack(p
))
5308 printk(KERN_INFO
"%-15.15s %c", p
->comm
, task_state_to_char(p
));
5310 if (p
->state
== TASK_RUNNING
)
5311 printk(KERN_CONT
" running task ");
5312 #ifdef CONFIG_DEBUG_STACK_USAGE
5313 free
= stack_not_used(p
);
5318 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5320 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5321 task_pid_nr(p
), ppid
,
5322 (unsigned long)task_thread_info(p
)->flags
);
5324 print_worker_info(KERN_INFO
, p
);
5325 show_stack(p
, NULL
);
5328 EXPORT_SYMBOL_GPL(sched_show_task
);
5331 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
5333 /* no filter, everything matches */
5337 /* filter, but doesn't match */
5338 if (!(p
->state
& state_filter
))
5342 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5345 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
5352 void show_state_filter(unsigned long state_filter
)
5354 struct task_struct
*g
, *p
;
5356 #if BITS_PER_LONG == 32
5358 " task PC stack pid father\n");
5361 " task PC stack pid father\n");
5364 for_each_process_thread(g
, p
) {
5366 * reset the NMI-timeout, listing all files on a slow
5367 * console might take a lot of time:
5368 * Also, reset softlockup watchdogs on all CPUs, because
5369 * another CPU might be blocked waiting for us to process
5372 touch_nmi_watchdog();
5373 touch_all_softlockup_watchdogs();
5374 if (state_filter_match(state_filter
, p
))
5378 #ifdef CONFIG_SCHED_DEBUG
5380 sysrq_sched_debug_show();
5384 * Only show locks if all tasks are dumped:
5387 debug_show_all_locks();
5391 * init_idle - set up an idle thread for a given CPU
5392 * @idle: task in question
5393 * @cpu: CPU the idle task belongs to
5395 * NOTE: this function does not set the idle thread's NEED_RESCHED
5396 * flag, to make booting more robust.
5398 void init_idle(struct task_struct
*idle
, int cpu
)
5400 struct rq
*rq
= cpu_rq(cpu
);
5401 unsigned long flags
;
5403 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5404 raw_spin_lock(&rq
->lock
);
5406 __sched_fork(0, idle
);
5407 idle
->state
= TASK_RUNNING
;
5408 idle
->se
.exec_start
= sched_clock();
5409 idle
->flags
|= PF_IDLE
;
5411 kasan_unpoison_task_stack(idle
);
5415 * Its possible that init_idle() gets called multiple times on a task,
5416 * in that case do_set_cpus_allowed() will not do the right thing.
5418 * And since this is boot we can forgo the serialization.
5420 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5423 * We're having a chicken and egg problem, even though we are
5424 * holding rq->lock, the CPU isn't yet set to this CPU so the
5425 * lockdep check in task_group() will fail.
5427 * Similar case to sched_fork(). / Alternatively we could
5428 * use task_rq_lock() here and obtain the other rq->lock.
5433 __set_task_cpu(idle
, cpu
);
5436 rq
->curr
= rq
->idle
= idle
;
5437 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5441 raw_spin_unlock(&rq
->lock
);
5442 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5444 /* Set the preempt count _outside_ the spinlocks! */
5445 init_idle_preempt_count(idle
, cpu
);
5448 * The idle tasks have their own, simple scheduling class:
5450 idle
->sched_class
= &idle_sched_class
;
5451 ftrace_graph_init_idle_task(idle
, cpu
);
5452 vtime_init_idle(idle
, cpu
);
5454 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5460 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5461 const struct cpumask
*trial
)
5465 if (!cpumask_weight(cur
))
5468 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
5473 int task_can_attach(struct task_struct
*p
,
5474 const struct cpumask
*cs_cpus_allowed
)
5479 * Kthreads which disallow setaffinity shouldn't be moved
5480 * to a new cpuset; we don't want to change their CPU
5481 * affinity and isolating such threads by their set of
5482 * allowed nodes is unnecessary. Thus, cpusets are not
5483 * applicable for such threads. This prevents checking for
5484 * success of set_cpus_allowed_ptr() on all attached tasks
5485 * before cpus_allowed may be changed.
5487 if (p
->flags
& PF_NO_SETAFFINITY
) {
5492 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5494 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
5500 bool sched_smp_initialized __read_mostly
;
5502 #ifdef CONFIG_NUMA_BALANCING
5503 /* Migrate current task p to target_cpu */
5504 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5506 struct migration_arg arg
= { p
, target_cpu
};
5507 int curr_cpu
= task_cpu(p
);
5509 if (curr_cpu
== target_cpu
)
5512 if (!cpumask_test_cpu(target_cpu
, &p
->cpus_allowed
))
5515 /* TODO: This is not properly updating schedstats */
5517 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5518 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5522 * Requeue a task on a given node and accurately track the number of NUMA
5523 * tasks on the runqueues
5525 void sched_setnuma(struct task_struct
*p
, int nid
)
5527 bool queued
, running
;
5531 rq
= task_rq_lock(p
, &rf
);
5532 queued
= task_on_rq_queued(p
);
5533 running
= task_current(rq
, p
);
5536 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5538 put_prev_task(rq
, p
);
5540 p
->numa_preferred_nid
= nid
;
5543 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
5545 set_curr_task(rq
, p
);
5546 task_rq_unlock(rq
, p
, &rf
);
5548 #endif /* CONFIG_NUMA_BALANCING */
5550 #ifdef CONFIG_HOTPLUG_CPU
5552 * Ensure that the idle task is using init_mm right before its CPU goes
5555 void idle_task_exit(void)
5557 struct mm_struct
*mm
= current
->active_mm
;
5559 BUG_ON(cpu_online(smp_processor_id()));
5561 if (mm
!= &init_mm
) {
5562 switch_mm(mm
, &init_mm
, current
);
5563 current
->active_mm
= &init_mm
;
5564 finish_arch_post_lock_switch();
5570 * Since this CPU is going 'away' for a while, fold any nr_active delta
5571 * we might have. Assumes we're called after migrate_tasks() so that the
5572 * nr_active count is stable. We need to take the teardown thread which
5573 * is calling this into account, so we hand in adjust = 1 to the load
5576 * Also see the comment "Global load-average calculations".
5578 static void calc_load_migrate(struct rq
*rq
)
5580 long delta
= calc_load_fold_active(rq
, 1);
5582 atomic_long_add(delta
, &calc_load_tasks
);
5585 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5589 static const struct sched_class fake_sched_class
= {
5590 .put_prev_task
= put_prev_task_fake
,
5593 static struct task_struct fake_task
= {
5595 * Avoid pull_{rt,dl}_task()
5597 .prio
= MAX_PRIO
+ 1,
5598 .sched_class
= &fake_sched_class
,
5602 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5603 * try_to_wake_up()->select_task_rq().
5605 * Called with rq->lock held even though we'er in stop_machine() and
5606 * there's no concurrency possible, we hold the required locks anyway
5607 * because of lock validation efforts.
5609 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
5611 struct rq
*rq
= dead_rq
;
5612 struct task_struct
*next
, *stop
= rq
->stop
;
5613 struct rq_flags orf
= *rf
;
5617 * Fudge the rq selection such that the below task selection loop
5618 * doesn't get stuck on the currently eligible stop task.
5620 * We're currently inside stop_machine() and the rq is either stuck
5621 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5622 * either way we should never end up calling schedule() until we're
5628 * put_prev_task() and pick_next_task() sched
5629 * class method both need to have an up-to-date
5630 * value of rq->clock[_task]
5632 update_rq_clock(rq
);
5636 * There's this thread running, bail when that's the only
5639 if (rq
->nr_running
== 1)
5643 * pick_next_task() assumes pinned rq->lock:
5645 next
= pick_next_task(rq
, &fake_task
, rf
);
5647 put_prev_task(rq
, next
);
5650 * Rules for changing task_struct::cpus_allowed are holding
5651 * both pi_lock and rq->lock, such that holding either
5652 * stabilizes the mask.
5654 * Drop rq->lock is not quite as disastrous as it usually is
5655 * because !cpu_active at this point, which means load-balance
5656 * will not interfere. Also, stop-machine.
5659 raw_spin_lock(&next
->pi_lock
);
5663 * Since we're inside stop-machine, _nothing_ should have
5664 * changed the task, WARN if weird stuff happened, because in
5665 * that case the above rq->lock drop is a fail too.
5667 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5668 raw_spin_unlock(&next
->pi_lock
);
5672 /* Find suitable destination for @next, with force if needed. */
5673 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5674 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
5675 if (rq
!= dead_rq
) {
5681 raw_spin_unlock(&next
->pi_lock
);
5686 #endif /* CONFIG_HOTPLUG_CPU */
5688 void set_rq_online(struct rq
*rq
)
5691 const struct sched_class
*class;
5693 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5696 for_each_class(class) {
5697 if (class->rq_online
)
5698 class->rq_online(rq
);
5703 void set_rq_offline(struct rq
*rq
)
5706 const struct sched_class
*class;
5708 for_each_class(class) {
5709 if (class->rq_offline
)
5710 class->rq_offline(rq
);
5713 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5718 static void set_cpu_rq_start_time(unsigned int cpu
)
5720 struct rq
*rq
= cpu_rq(cpu
);
5722 rq
->age_stamp
= sched_clock_cpu(cpu
);
5726 * used to mark begin/end of suspend/resume:
5728 static int num_cpus_frozen
;
5731 * Update cpusets according to cpu_active mask. If cpusets are
5732 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5733 * around partition_sched_domains().
5735 * If we come here as part of a suspend/resume, don't touch cpusets because we
5736 * want to restore it back to its original state upon resume anyway.
5738 static void cpuset_cpu_active(void)
5740 if (cpuhp_tasks_frozen
) {
5742 * num_cpus_frozen tracks how many CPUs are involved in suspend
5743 * resume sequence. As long as this is not the last online
5744 * operation in the resume sequence, just build a single sched
5745 * domain, ignoring cpusets.
5747 partition_sched_domains(1, NULL
, NULL
);
5748 if (--num_cpus_frozen
)
5751 * This is the last CPU online operation. So fall through and
5752 * restore the original sched domains by considering the
5753 * cpuset configurations.
5755 cpuset_force_rebuild();
5757 cpuset_update_active_cpus();
5760 static int cpuset_cpu_inactive(unsigned int cpu
)
5762 if (!cpuhp_tasks_frozen
) {
5763 if (dl_cpu_busy(cpu
))
5765 cpuset_update_active_cpus();
5768 partition_sched_domains(1, NULL
, NULL
);
5773 int sched_cpu_activate(unsigned int cpu
)
5775 struct rq
*rq
= cpu_rq(cpu
);
5778 set_cpu_active(cpu
, true);
5780 if (sched_smp_initialized
) {
5781 sched_domains_numa_masks_set(cpu
);
5782 cpuset_cpu_active();
5786 * Put the rq online, if not already. This happens:
5788 * 1) In the early boot process, because we build the real domains
5789 * after all CPUs have been brought up.
5791 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5794 rq_lock_irqsave(rq
, &rf
);
5796 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5799 rq_unlock_irqrestore(rq
, &rf
);
5801 update_max_interval();
5806 int sched_cpu_deactivate(unsigned int cpu
)
5810 set_cpu_active(cpu
, false);
5812 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5813 * users of this state to go away such that all new such users will
5816 * Do sync before park smpboot threads to take care the rcu boost case.
5818 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
5820 if (!sched_smp_initialized
)
5823 ret
= cpuset_cpu_inactive(cpu
);
5825 set_cpu_active(cpu
, true);
5828 sched_domains_numa_masks_clear(cpu
);
5832 static void sched_rq_cpu_starting(unsigned int cpu
)
5834 struct rq
*rq
= cpu_rq(cpu
);
5836 rq
->calc_load_update
= calc_load_update
;
5837 update_max_interval();
5840 int sched_cpu_starting(unsigned int cpu
)
5842 set_cpu_rq_start_time(cpu
);
5843 sched_rq_cpu_starting(cpu
);
5844 sched_tick_start(cpu
);
5848 #ifdef CONFIG_HOTPLUG_CPU
5849 int sched_cpu_dying(unsigned int cpu
)
5851 struct rq
*rq
= cpu_rq(cpu
);
5854 /* Handle pending wakeups and then migrate everything off */
5855 sched_ttwu_pending();
5856 sched_tick_stop(cpu
);
5858 rq_lock_irqsave(rq
, &rf
);
5860 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5863 migrate_tasks(rq
, &rf
);
5864 BUG_ON(rq
->nr_running
!= 1);
5865 rq_unlock_irqrestore(rq
, &rf
);
5867 calc_load_migrate(rq
);
5868 update_max_interval();
5869 nohz_balance_exit_idle(rq
);
5875 #ifdef CONFIG_SCHED_SMT
5876 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
5878 static void sched_init_smt(void)
5881 * We've enumerated all CPUs and will assume that if any CPU
5882 * has SMT siblings, CPU0 will too.
5884 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5885 static_branch_enable(&sched_smt_present
);
5888 static inline void sched_init_smt(void) { }
5891 void __init
sched_init_smp(void)
5896 * There's no userspace yet to cause hotplug operations; hence all the
5897 * CPU masks are stable and all blatant races in the below code cannot
5900 mutex_lock(&sched_domains_mutex
);
5901 sched_init_domains(cpu_active_mask
);
5902 mutex_unlock(&sched_domains_mutex
);
5904 /* Move init over to a non-isolated CPU */
5905 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
5907 sched_init_granularity();
5909 init_sched_rt_class();
5910 init_sched_dl_class();
5914 sched_smp_initialized
= true;
5917 static int __init
migration_init(void)
5919 sched_rq_cpu_starting(smp_processor_id());
5922 early_initcall(migration_init
);
5925 void __init
sched_init_smp(void)
5927 sched_init_granularity();
5929 #endif /* CONFIG_SMP */
5931 int in_sched_functions(unsigned long addr
)
5933 return in_lock_functions(addr
) ||
5934 (addr
>= (unsigned long)__sched_text_start
5935 && addr
< (unsigned long)__sched_text_end
);
5938 #ifdef CONFIG_CGROUP_SCHED
5940 * Default task group.
5941 * Every task in system belongs to this group at bootup.
5943 struct task_group root_task_group
;
5944 LIST_HEAD(task_groups
);
5946 /* Cacheline aligned slab cache for task_group */
5947 static struct kmem_cache
*task_group_cache __read_mostly
;
5950 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5951 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5953 void __init
sched_init(void)
5956 unsigned long alloc_size
= 0, ptr
;
5961 #ifdef CONFIG_FAIR_GROUP_SCHED
5962 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5964 #ifdef CONFIG_RT_GROUP_SCHED
5965 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5968 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
5970 #ifdef CONFIG_FAIR_GROUP_SCHED
5971 root_task_group
.se
= (struct sched_entity
**)ptr
;
5972 ptr
+= nr_cpu_ids
* sizeof(void **);
5974 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
5975 ptr
+= nr_cpu_ids
* sizeof(void **);
5977 #endif /* CONFIG_FAIR_GROUP_SCHED */
5978 #ifdef CONFIG_RT_GROUP_SCHED
5979 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
5980 ptr
+= nr_cpu_ids
* sizeof(void **);
5982 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
5983 ptr
+= nr_cpu_ids
* sizeof(void **);
5985 #endif /* CONFIG_RT_GROUP_SCHED */
5987 #ifdef CONFIG_CPUMASK_OFFSTACK
5988 for_each_possible_cpu(i
) {
5989 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5990 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5991 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5992 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5994 #endif /* CONFIG_CPUMASK_OFFSTACK */
5996 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
5997 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
6000 init_defrootdomain();
6003 #ifdef CONFIG_RT_GROUP_SCHED
6004 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6005 global_rt_period(), global_rt_runtime());
6006 #endif /* CONFIG_RT_GROUP_SCHED */
6008 #ifdef CONFIG_CGROUP_SCHED
6009 task_group_cache
= KMEM_CACHE(task_group
, 0);
6011 list_add(&root_task_group
.list
, &task_groups
);
6012 INIT_LIST_HEAD(&root_task_group
.children
);
6013 INIT_LIST_HEAD(&root_task_group
.siblings
);
6014 autogroup_init(&init_task
);
6015 #endif /* CONFIG_CGROUP_SCHED */
6017 for_each_possible_cpu(i
) {
6021 raw_spin_lock_init(&rq
->lock
);
6023 rq
->calc_load_active
= 0;
6024 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6025 init_cfs_rq(&rq
->cfs
);
6026 init_rt_rq(&rq
->rt
);
6027 init_dl_rq(&rq
->dl
);
6028 #ifdef CONFIG_FAIR_GROUP_SCHED
6029 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6030 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6031 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
6033 * How much CPU bandwidth does root_task_group get?
6035 * In case of task-groups formed thr' the cgroup filesystem, it
6036 * gets 100% of the CPU resources in the system. This overall
6037 * system CPU resource is divided among the tasks of
6038 * root_task_group and its child task-groups in a fair manner,
6039 * based on each entity's (task or task-group's) weight
6040 * (se->load.weight).
6042 * In other words, if root_task_group has 10 tasks of weight
6043 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6044 * then A0's share of the CPU resource is:
6046 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6048 * We achieve this by letting root_task_group's tasks sit
6049 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6051 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6052 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6053 #endif /* CONFIG_FAIR_GROUP_SCHED */
6055 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6056 #ifdef CONFIG_RT_GROUP_SCHED
6057 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6060 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6061 rq
->cpu_load
[j
] = 0;
6066 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
6067 rq
->balance_callback
= NULL
;
6068 rq
->active_balance
= 0;
6069 rq
->next_balance
= jiffies
;
6074 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6075 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6077 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6079 rq_attach_root(rq
, &def_root_domain
);
6080 #ifdef CONFIG_NO_HZ_COMMON
6081 rq
->last_load_update_tick
= jiffies
;
6082 rq
->last_blocked_load_update_tick
= jiffies
;
6083 atomic_set(&rq
->nohz_flags
, 0);
6085 #endif /* CONFIG_SMP */
6087 atomic_set(&rq
->nr_iowait
, 0);
6090 set_load_weight(&init_task
, false);
6093 * The boot idle thread does lazy MMU switching as well:
6096 enter_lazy_tlb(&init_mm
, current
);
6099 * Make us the idle thread. Technically, schedule() should not be
6100 * called from this thread, however somewhere below it might be,
6101 * but because we are the idle thread, we just pick up running again
6102 * when this runqueue becomes "idle".
6104 init_idle(current
, smp_processor_id());
6106 calc_load_update
= jiffies
+ LOAD_FREQ
;
6109 idle_thread_set_boot_cpu();
6110 set_cpu_rq_start_time(smp_processor_id());
6112 init_sched_fair_class();
6116 scheduler_running
= 1;
6119 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6120 static inline int preempt_count_equals(int preempt_offset
)
6122 int nested
= preempt_count() + rcu_preempt_depth();
6124 return (nested
== preempt_offset
);
6127 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6130 * Blocking primitives will set (and therefore destroy) current->state,
6131 * since we will exit with TASK_RUNNING make sure we enter with it,
6132 * otherwise we will destroy state.
6134 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6135 "do not call blocking ops when !TASK_RUNNING; "
6136 "state=%lx set at [<%p>] %pS\n",
6138 (void *)current
->task_state_change
,
6139 (void *)current
->task_state_change
);
6141 ___might_sleep(file
, line
, preempt_offset
);
6143 EXPORT_SYMBOL(__might_sleep
);
6145 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6147 /* Ratelimiting timestamp: */
6148 static unsigned long prev_jiffy
;
6150 unsigned long preempt_disable_ip
;
6152 /* WARN_ON_ONCE() by default, no rate limit required: */
6155 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6156 !is_idle_task(current
)) ||
6157 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
6161 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6163 prev_jiffy
= jiffies
;
6165 /* Save this before calling printk(), since that will clobber it: */
6166 preempt_disable_ip
= get_preempt_disable_ip(current
);
6169 "BUG: sleeping function called from invalid context at %s:%d\n",
6172 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6173 in_atomic(), irqs_disabled(),
6174 current
->pid
, current
->comm
);
6176 if (task_stack_end_corrupted(current
))
6177 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6179 debug_show_held_locks(current
);
6180 if (irqs_disabled())
6181 print_irqtrace_events(current
);
6182 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6183 && !preempt_count_equals(preempt_offset
)) {
6184 pr_err("Preemption disabled at:");
6185 print_ip_sym(preempt_disable_ip
);
6189 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6191 EXPORT_SYMBOL(___might_sleep
);
6194 #ifdef CONFIG_MAGIC_SYSRQ
6195 void normalize_rt_tasks(void)
6197 struct task_struct
*g
, *p
;
6198 struct sched_attr attr
= {
6199 .sched_policy
= SCHED_NORMAL
,
6202 read_lock(&tasklist_lock
);
6203 for_each_process_thread(g
, p
) {
6205 * Only normalize user tasks:
6207 if (p
->flags
& PF_KTHREAD
)
6210 p
->se
.exec_start
= 0;
6211 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6212 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6213 schedstat_set(p
->se
.statistics
.block_start
, 0);
6215 if (!dl_task(p
) && !rt_task(p
)) {
6217 * Renice negative nice level userspace
6220 if (task_nice(p
) < 0)
6221 set_user_nice(p
, 0);
6225 __sched_setscheduler(p
, &attr
, false, false);
6227 read_unlock(&tasklist_lock
);
6230 #endif /* CONFIG_MAGIC_SYSRQ */
6232 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6234 * These functions are only useful for the IA64 MCA handling, or kdb.
6236 * They can only be called when the whole system has been
6237 * stopped - every CPU needs to be quiescent, and no scheduling
6238 * activity can take place. Using them for anything else would
6239 * be a serious bug, and as a result, they aren't even visible
6240 * under any other configuration.
6244 * curr_task - return the current task for a given CPU.
6245 * @cpu: the processor in question.
6247 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6249 * Return: The current task for @cpu.
6251 struct task_struct
*curr_task(int cpu
)
6253 return cpu_curr(cpu
);
6256 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6260 * set_curr_task - set the current task for a given CPU.
6261 * @cpu: the processor in question.
6262 * @p: the task pointer to set.
6264 * Description: This function must only be used when non-maskable interrupts
6265 * are serviced on a separate stack. It allows the architecture to switch the
6266 * notion of the current task on a CPU in a non-blocking manner. This function
6267 * must be called with all CPU's synchronized, and interrupts disabled, the
6268 * and caller must save the original value of the current task (see
6269 * curr_task() above) and restore that value before reenabling interrupts and
6270 * re-starting the system.
6272 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6274 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6281 #ifdef CONFIG_CGROUP_SCHED
6282 /* task_group_lock serializes the addition/removal of task groups */
6283 static DEFINE_SPINLOCK(task_group_lock
);
6285 static void sched_free_group(struct task_group
*tg
)
6287 free_fair_sched_group(tg
);
6288 free_rt_sched_group(tg
);
6290 kmem_cache_free(task_group_cache
, tg
);
6293 /* allocate runqueue etc for a new task group */
6294 struct task_group
*sched_create_group(struct task_group
*parent
)
6296 struct task_group
*tg
;
6298 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
6300 return ERR_PTR(-ENOMEM
);
6302 if (!alloc_fair_sched_group(tg
, parent
))
6305 if (!alloc_rt_sched_group(tg
, parent
))
6311 sched_free_group(tg
);
6312 return ERR_PTR(-ENOMEM
);
6315 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6317 unsigned long flags
;
6319 spin_lock_irqsave(&task_group_lock
, flags
);
6320 list_add_rcu(&tg
->list
, &task_groups
);
6322 /* Root should already exist: */
6325 tg
->parent
= parent
;
6326 INIT_LIST_HEAD(&tg
->children
);
6327 list_add_rcu(&tg
->siblings
, &parent
->children
);
6328 spin_unlock_irqrestore(&task_group_lock
, flags
);
6330 online_fair_sched_group(tg
);
6333 /* rcu callback to free various structures associated with a task group */
6334 static void sched_free_group_rcu(struct rcu_head
*rhp
)
6336 /* Now it should be safe to free those cfs_rqs: */
6337 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
6340 void sched_destroy_group(struct task_group
*tg
)
6342 /* Wait for possible concurrent references to cfs_rqs complete: */
6343 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
6346 void sched_offline_group(struct task_group
*tg
)
6348 unsigned long flags
;
6350 /* End participation in shares distribution: */
6351 unregister_fair_sched_group(tg
);
6353 spin_lock_irqsave(&task_group_lock
, flags
);
6354 list_del_rcu(&tg
->list
);
6355 list_del_rcu(&tg
->siblings
);
6356 spin_unlock_irqrestore(&task_group_lock
, flags
);
6359 static void sched_change_group(struct task_struct
*tsk
, int type
)
6361 struct task_group
*tg
;
6364 * All callers are synchronized by task_rq_lock(); we do not use RCU
6365 * which is pointless here. Thus, we pass "true" to task_css_check()
6366 * to prevent lockdep warnings.
6368 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
6369 struct task_group
, css
);
6370 tg
= autogroup_task_group(tsk
, tg
);
6371 tsk
->sched_task_group
= tg
;
6373 #ifdef CONFIG_FAIR_GROUP_SCHED
6374 if (tsk
->sched_class
->task_change_group
)
6375 tsk
->sched_class
->task_change_group(tsk
, type
);
6378 set_task_rq(tsk
, task_cpu(tsk
));
6382 * Change task's runqueue when it moves between groups.
6384 * The caller of this function should have put the task in its new group by
6385 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6388 void sched_move_task(struct task_struct
*tsk
)
6390 int queued
, running
, queue_flags
=
6391 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
6395 rq
= task_rq_lock(tsk
, &rf
);
6396 update_rq_clock(rq
);
6398 running
= task_current(rq
, tsk
);
6399 queued
= task_on_rq_queued(tsk
);
6402 dequeue_task(rq
, tsk
, queue_flags
);
6404 put_prev_task(rq
, tsk
);
6406 sched_change_group(tsk
, TASK_MOVE_GROUP
);
6409 enqueue_task(rq
, tsk
, queue_flags
);
6411 set_curr_task(rq
, tsk
);
6413 task_rq_unlock(rq
, tsk
, &rf
);
6416 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
6418 return css
? container_of(css
, struct task_group
, css
) : NULL
;
6421 static struct cgroup_subsys_state
*
6422 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6424 struct task_group
*parent
= css_tg(parent_css
);
6425 struct task_group
*tg
;
6428 /* This is early initialization for the top cgroup */
6429 return &root_task_group
.css
;
6432 tg
= sched_create_group(parent
);
6434 return ERR_PTR(-ENOMEM
);
6439 /* Expose task group only after completing cgroup initialization */
6440 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
6442 struct task_group
*tg
= css_tg(css
);
6443 struct task_group
*parent
= css_tg(css
->parent
);
6446 sched_online_group(tg
, parent
);
6450 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
6452 struct task_group
*tg
= css_tg(css
);
6454 sched_offline_group(tg
);
6457 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
6459 struct task_group
*tg
= css_tg(css
);
6462 * Relies on the RCU grace period between css_released() and this.
6464 sched_free_group(tg
);
6468 * This is called before wake_up_new_task(), therefore we really only
6469 * have to set its group bits, all the other stuff does not apply.
6471 static void cpu_cgroup_fork(struct task_struct
*task
)
6476 rq
= task_rq_lock(task
, &rf
);
6478 update_rq_clock(rq
);
6479 sched_change_group(task
, TASK_SET_GROUP
);
6481 task_rq_unlock(rq
, task
, &rf
);
6484 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
6486 struct task_struct
*task
;
6487 struct cgroup_subsys_state
*css
;
6490 cgroup_taskset_for_each(task
, css
, tset
) {
6491 #ifdef CONFIG_RT_GROUP_SCHED
6492 if (!sched_rt_can_attach(css_tg(css
), task
))
6495 /* We don't support RT-tasks being in separate groups */
6496 if (task
->sched_class
!= &fair_sched_class
)
6500 * Serialize against wake_up_new_task() such that if its
6501 * running, we're sure to observe its full state.
6503 raw_spin_lock_irq(&task
->pi_lock
);
6505 * Avoid calling sched_move_task() before wake_up_new_task()
6506 * has happened. This would lead to problems with PELT, due to
6507 * move wanting to detach+attach while we're not attached yet.
6509 if (task
->state
== TASK_NEW
)
6511 raw_spin_unlock_irq(&task
->pi_lock
);
6519 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
6521 struct task_struct
*task
;
6522 struct cgroup_subsys_state
*css
;
6524 cgroup_taskset_for_each(task
, css
, tset
)
6525 sched_move_task(task
);
6528 #ifdef CONFIG_FAIR_GROUP_SCHED
6529 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
6530 struct cftype
*cftype
, u64 shareval
)
6532 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
6535 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
6538 struct task_group
*tg
= css_tg(css
);
6540 return (u64
) scale_load_down(tg
->shares
);
6543 #ifdef CONFIG_CFS_BANDWIDTH
6544 static DEFINE_MUTEX(cfs_constraints_mutex
);
6546 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
6547 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
6549 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
6551 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
6553 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
6554 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6556 if (tg
== &root_task_group
)
6560 * Ensure we have at some amount of bandwidth every period. This is
6561 * to prevent reaching a state of large arrears when throttled via
6562 * entity_tick() resulting in prolonged exit starvation.
6564 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
6568 * Likewise, bound things on the otherside by preventing insane quota
6569 * periods. This also allows us to normalize in computing quota
6572 if (period
> max_cfs_quota_period
)
6576 * Prevent race between setting of cfs_rq->runtime_enabled and
6577 * unthrottle_offline_cfs_rqs().
6580 mutex_lock(&cfs_constraints_mutex
);
6581 ret
= __cfs_schedulable(tg
, period
, quota
);
6585 runtime_enabled
= quota
!= RUNTIME_INF
;
6586 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
6588 * If we need to toggle cfs_bandwidth_used, off->on must occur
6589 * before making related changes, and on->off must occur afterwards
6591 if (runtime_enabled
&& !runtime_was_enabled
)
6592 cfs_bandwidth_usage_inc();
6593 raw_spin_lock_irq(&cfs_b
->lock
);
6594 cfs_b
->period
= ns_to_ktime(period
);
6595 cfs_b
->quota
= quota
;
6597 __refill_cfs_bandwidth_runtime(cfs_b
);
6599 /* Restart the period timer (if active) to handle new period expiry: */
6600 if (runtime_enabled
)
6601 start_cfs_bandwidth(cfs_b
);
6603 raw_spin_unlock_irq(&cfs_b
->lock
);
6605 for_each_online_cpu(i
) {
6606 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
6607 struct rq
*rq
= cfs_rq
->rq
;
6610 rq_lock_irq(rq
, &rf
);
6611 cfs_rq
->runtime_enabled
= runtime_enabled
;
6612 cfs_rq
->runtime_remaining
= 0;
6614 if (cfs_rq
->throttled
)
6615 unthrottle_cfs_rq(cfs_rq
);
6616 rq_unlock_irq(rq
, &rf
);
6618 if (runtime_was_enabled
&& !runtime_enabled
)
6619 cfs_bandwidth_usage_dec();
6621 mutex_unlock(&cfs_constraints_mutex
);
6627 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
6631 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
6632 if (cfs_quota_us
< 0)
6633 quota
= RUNTIME_INF
;
6635 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
6637 return tg_set_cfs_bandwidth(tg
, period
, quota
);
6640 long tg_get_cfs_quota(struct task_group
*tg
)
6644 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
6647 quota_us
= tg
->cfs_bandwidth
.quota
;
6648 do_div(quota_us
, NSEC_PER_USEC
);
6653 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
6657 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
6658 quota
= tg
->cfs_bandwidth
.quota
;
6660 return tg_set_cfs_bandwidth(tg
, period
, quota
);
6663 long tg_get_cfs_period(struct task_group
*tg
)
6667 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
6668 do_div(cfs_period_us
, NSEC_PER_USEC
);
6670 return cfs_period_us
;
6673 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
6676 return tg_get_cfs_quota(css_tg(css
));
6679 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
6680 struct cftype
*cftype
, s64 cfs_quota_us
)
6682 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
6685 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
6688 return tg_get_cfs_period(css_tg(css
));
6691 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
6692 struct cftype
*cftype
, u64 cfs_period_us
)
6694 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
6697 struct cfs_schedulable_data
{
6698 struct task_group
*tg
;
6703 * normalize group quota/period to be quota/max_period
6704 * note: units are usecs
6706 static u64
normalize_cfs_quota(struct task_group
*tg
,
6707 struct cfs_schedulable_data
*d
)
6715 period
= tg_get_cfs_period(tg
);
6716 quota
= tg_get_cfs_quota(tg
);
6719 /* note: these should typically be equivalent */
6720 if (quota
== RUNTIME_INF
|| quota
== -1)
6723 return to_ratio(period
, quota
);
6726 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
6728 struct cfs_schedulable_data
*d
= data
;
6729 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6730 s64 quota
= 0, parent_quota
= -1;
6733 quota
= RUNTIME_INF
;
6735 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
6737 quota
= normalize_cfs_quota(tg
, d
);
6738 parent_quota
= parent_b
->hierarchical_quota
;
6741 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6742 * always take the min. On cgroup1, only inherit when no
6745 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
6746 quota
= min(quota
, parent_quota
);
6748 if (quota
== RUNTIME_INF
)
6749 quota
= parent_quota
;
6750 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
6754 cfs_b
->hierarchical_quota
= quota
;
6759 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
6762 struct cfs_schedulable_data data
= {
6768 if (quota
!= RUNTIME_INF
) {
6769 do_div(data
.period
, NSEC_PER_USEC
);
6770 do_div(data
.quota
, NSEC_PER_USEC
);
6774 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
6780 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
6782 struct task_group
*tg
= css_tg(seq_css(sf
));
6783 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6785 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
6786 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
6787 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
6791 #endif /* CONFIG_CFS_BANDWIDTH */
6792 #endif /* CONFIG_FAIR_GROUP_SCHED */
6794 #ifdef CONFIG_RT_GROUP_SCHED
6795 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
6796 struct cftype
*cft
, s64 val
)
6798 return sched_group_set_rt_runtime(css_tg(css
), val
);
6801 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
6804 return sched_group_rt_runtime(css_tg(css
));
6807 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
6808 struct cftype
*cftype
, u64 rt_period_us
)
6810 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
6813 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
6816 return sched_group_rt_period(css_tg(css
));
6818 #endif /* CONFIG_RT_GROUP_SCHED */
6820 static struct cftype cpu_legacy_files
[] = {
6821 #ifdef CONFIG_FAIR_GROUP_SCHED
6824 .read_u64
= cpu_shares_read_u64
,
6825 .write_u64
= cpu_shares_write_u64
,
6828 #ifdef CONFIG_CFS_BANDWIDTH
6830 .name
= "cfs_quota_us",
6831 .read_s64
= cpu_cfs_quota_read_s64
,
6832 .write_s64
= cpu_cfs_quota_write_s64
,
6835 .name
= "cfs_period_us",
6836 .read_u64
= cpu_cfs_period_read_u64
,
6837 .write_u64
= cpu_cfs_period_write_u64
,
6841 .seq_show
= cpu_cfs_stat_show
,
6844 #ifdef CONFIG_RT_GROUP_SCHED
6846 .name
= "rt_runtime_us",
6847 .read_s64
= cpu_rt_runtime_read
,
6848 .write_s64
= cpu_rt_runtime_write
,
6851 .name
= "rt_period_us",
6852 .read_u64
= cpu_rt_period_read_uint
,
6853 .write_u64
= cpu_rt_period_write_uint
,
6859 static int cpu_extra_stat_show(struct seq_file
*sf
,
6860 struct cgroup_subsys_state
*css
)
6862 #ifdef CONFIG_CFS_BANDWIDTH
6864 struct task_group
*tg
= css_tg(css
);
6865 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6868 throttled_usec
= cfs_b
->throttled_time
;
6869 do_div(throttled_usec
, NSEC_PER_USEC
);
6871 seq_printf(sf
, "nr_periods %d\n"
6873 "throttled_usec %llu\n",
6874 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
6881 #ifdef CONFIG_FAIR_GROUP_SCHED
6882 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
6885 struct task_group
*tg
= css_tg(css
);
6886 u64 weight
= scale_load_down(tg
->shares
);
6888 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
6891 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
6892 struct cftype
*cft
, u64 weight
)
6895 * cgroup weight knobs should use the common MIN, DFL and MAX
6896 * values which are 1, 100 and 10000 respectively. While it loses
6897 * a bit of range on both ends, it maps pretty well onto the shares
6898 * value used by scheduler and the round-trip conversions preserve
6899 * the original value over the entire range.
6901 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
6904 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
6906 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
6909 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
6912 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
6913 int last_delta
= INT_MAX
;
6916 /* find the closest nice value to the current weight */
6917 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
6918 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
6919 if (delta
>= last_delta
)
6924 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
6927 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
6928 struct cftype
*cft
, s64 nice
)
6930 unsigned long weight
;
6932 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
6935 weight
= sched_prio_to_weight
[NICE_TO_PRIO(nice
) - MAX_RT_PRIO
];
6936 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
6940 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
6941 long period
, long quota
)
6944 seq_puts(sf
, "max");
6946 seq_printf(sf
, "%ld", quota
);
6948 seq_printf(sf
, " %ld\n", period
);
6951 /* caller should put the current value in *@periodp before calling */
6952 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
6953 u64
*periodp
, u64
*quotap
)
6955 char tok
[21]; /* U64_MAX */
6957 if (!sscanf(buf
, "%s %llu", tok
, periodp
))
6960 *periodp
*= NSEC_PER_USEC
;
6962 if (sscanf(tok
, "%llu", quotap
))
6963 *quotap
*= NSEC_PER_USEC
;
6964 else if (!strcmp(tok
, "max"))
6965 *quotap
= RUNTIME_INF
;
6972 #ifdef CONFIG_CFS_BANDWIDTH
6973 static int cpu_max_show(struct seq_file
*sf
, void *v
)
6975 struct task_group
*tg
= css_tg(seq_css(sf
));
6977 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
6981 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
6982 char *buf
, size_t nbytes
, loff_t off
)
6984 struct task_group
*tg
= css_tg(of_css(of
));
6985 u64 period
= tg_get_cfs_period(tg
);
6989 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
6991 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
6992 return ret
?: nbytes
;
6996 static struct cftype cpu_files
[] = {
6997 #ifdef CONFIG_FAIR_GROUP_SCHED
7000 .flags
= CFTYPE_NOT_ON_ROOT
,
7001 .read_u64
= cpu_weight_read_u64
,
7002 .write_u64
= cpu_weight_write_u64
,
7005 .name
= "weight.nice",
7006 .flags
= CFTYPE_NOT_ON_ROOT
,
7007 .read_s64
= cpu_weight_nice_read_s64
,
7008 .write_s64
= cpu_weight_nice_write_s64
,
7011 #ifdef CONFIG_CFS_BANDWIDTH
7014 .flags
= CFTYPE_NOT_ON_ROOT
,
7015 .seq_show
= cpu_max_show
,
7016 .write
= cpu_max_write
,
7022 struct cgroup_subsys cpu_cgrp_subsys
= {
7023 .css_alloc
= cpu_cgroup_css_alloc
,
7024 .css_online
= cpu_cgroup_css_online
,
7025 .css_released
= cpu_cgroup_css_released
,
7026 .css_free
= cpu_cgroup_css_free
,
7027 .css_extra_stat_show
= cpu_extra_stat_show
,
7028 .fork
= cpu_cgroup_fork
,
7029 .can_attach
= cpu_cgroup_can_attach
,
7030 .attach
= cpu_cgroup_attach
,
7031 .legacy_cftypes
= cpu_legacy_files
,
7032 .dfl_cftypes
= cpu_files
,
7037 #endif /* CONFIG_CGROUP_SCHED */
7039 void dump_cpu_task(int cpu
)
7041 pr_info("Task dump for CPU %d:\n", cpu
);
7042 sched_show_task(cpu_curr(cpu
));
7046 * Nice levels are multiplicative, with a gentle 10% change for every
7047 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7048 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7049 * that remained on nice 0.
7051 * The "10% effect" is relative and cumulative: from _any_ nice level,
7052 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7053 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7054 * If a task goes up by ~10% and another task goes down by ~10% then
7055 * the relative distance between them is ~25%.)
7057 const int sched_prio_to_weight
[40] = {
7058 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7059 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7060 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7061 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7062 /* 0 */ 1024, 820, 655, 526, 423,
7063 /* 5 */ 335, 272, 215, 172, 137,
7064 /* 10 */ 110, 87, 70, 56, 45,
7065 /* 15 */ 36, 29, 23, 18, 15,
7069 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7071 * In cases where the weight does not change often, we can use the
7072 * precalculated inverse to speed up arithmetics by turning divisions
7073 * into multiplications:
7075 const u32 sched_prio_to_wmult
[40] = {
7076 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7077 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7078 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7079 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7080 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7081 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7082 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7083 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7086 #undef CREATE_TRACE_POINTS